Add knowledge Q&A API: cited answers, document search, and store stats

app/common/src/main/java/stirling/software/SPDF/config/EndpointConfiguration.java

@@ -433,9 +433,19 @@ public class EndpointConfiguration {
         addEndpointToGroup("Automation", "automate"); // Alias for handleData (user-friendly name)
         addEndpointToGroup("Automation", "pipeline");
 
-        // Adding endpoints to "DocParse" group (ingestion: chunk + index + export)
+        // Adding endpoints to "DocParse" group (parsing, splitting, chunking, extraction,
+        // templating)
+        addEndpointToGroup("DocParse", "parse-document");
+        addEndpointToGroup("DocParse", "extract-fields");
+        addEndpointToGroup("DocParse", "smart-split");
+        addEndpointToGroup("DocParse", "chunk-document");
         addEndpointToGroup("DocParse", "rag-ingest");
+        addEndpointToGroup("DocParse", "rag-documents");
+        addEndpointToGroup("DocParse", "rag-search");
+        addEndpointToGroup("DocParse", "rag-ask");
         addEndpointToGroup("DocParse", "extract-tables");
+        addEndpointToGroup("DocParse", "suggest-schema");
+        addEndpointToGroup("DocParse", "fill-template");
 
         // Adding endpoints to "DeveloperTools" group
         addEndpointToGroup("DeveloperTools", "show-javascript");

app/core/src/main/java/stirling/software/SPDF/controller/api/misc/ConfigController.java

@@ -24,6 +24,7 @@ import stirling.software.common.annotations.api.ConfigApi;
 import stirling.software.common.configuration.AppConfig;
 import stirling.software.common.configuration.interfaces.ShowAdminInterface;
 import stirling.software.common.model.ApplicationProperties;
+import stirling.software.common.service.DocparseCapabilityServiceInterface;
 import stirling.software.common.service.ServerCertificateServiceInterface;
 import stirling.software.common.service.UserServiceInterface;
 import stirling.software.common.util.GeneralUtils;
@@ -41,6 +42,7 @@ public class ConfigController {
     private final ShowAdminInterface showAdmin;
     private final stirling.software.common.service.LicenseServiceInterface licenseService;
     private final stirling.software.SPDF.config.ExternalAppDepConfig externalAppDepConfig;
+    private final DocparseCapabilityServiceInterface docparseCapabilityService;
 
     public ConfigController(
             ApplicationProperties applicationProperties,
@@ -54,7 +56,9 @@ public class ConfigController {
                     ShowAdminInterface showAdmin,
             @org.springframework.beans.factory.annotation.Autowired(required = false)
                     stirling.software.common.service.LicenseServiceInterface licenseService,
-            stirling.software.SPDF.config.ExternalAppDepConfig externalAppDepConfig) {
+            stirling.software.SPDF.config.ExternalAppDepConfig externalAppDepConfig,
+            @org.springframework.beans.factory.annotation.Autowired(required = false)
+                    DocparseCapabilityServiceInterface docparseCapabilityService) {
         this.applicationProperties = applicationProperties;
         this.applicationContext = applicationContext;
         this.endpointConfiguration = endpointConfiguration;
@@ -63,6 +67,7 @@ public class ConfigController {
         this.showAdmin = showAdmin;
         this.licenseService = licenseService;
         this.externalAppDepConfig = externalAppDepConfig;
+        this.docparseCapabilityService = docparseCapabilityService;
     }
 
     /**
@@ -350,6 +355,16 @@ public class ConfigController {
                             Map.entry("pdfComment", aiFeatures.isPdfComment()),
                             Map.entry("classify", aiFeatures.isClassify())));
 
+            // DocParse settings; "advanced" reflects the cached engine capability probe and is
+            // false when the engine is disabled, unreachable, or the proprietary module is absent.
+            boolean docparseEnabled = applicationProperties.getDocparse().isEnabled();
+            configData.put("docparseEnabled", docparseEnabled);
+            configData.put(
+                    "docparseAdvanced",
+                    docparseEnabled
+                            && docparseCapabilityService != null
+                            && docparseCapabilityService.isAdvancedInstalled());
+
             // Timestamp TSA settings — single source of truth for presets + admin URLs
             ApplicationProperties.Security.Timestamp tsConfig =
                     applicationProperties.getSecurity().getTimestamp();

app/core/src/test/java/stirling/software/SPDF/controller/api/misc/ConfigControllerMoreTest.java

@@ -74,7 +74,8 @@ class ConfigControllerMoreTest {
                 userService,
                 showAdmin,
                 licenseService,
-                externalAppDepConfig);
+                externalAppDepConfig,
+                null);
     }
 
     @SuppressWarnings("unchecked")

app/core/src/test/java/stirling/software/SPDF/controller/api/misc/ConfigControllerTest.java

@@ -52,7 +52,8 @@ class ConfigControllerTest {
                         userService,
                         showAdmin,
                         licenseService,
-                        mock(stirling.software.SPDF.config.ExternalAppDepConfig.class));
+                        mock(stirling.software.SPDF.config.ExternalAppDepConfig.class),
+                        null);
     }
 
     @Test

app/proprietary/src/main/java/stirling/software/proprietary/controller/api/DocParseController.java

@@ -4,22 +4,32 @@ import java.io.ByteArrayOutputStream;
 import java.io.IOException;
 import java.io.StringWriter;
 import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.StandardCopyOption;
+import java.util.Base64;
 import java.util.List;
+import java.util.Locale;
 import java.util.zip.ZipEntry;
 import java.util.zip.ZipOutputStream;
 
 import org.apache.commons.csv.CSVFormat;
 import org.apache.commons.csv.CSVPrinter;
+import org.apache.pdfbox.pdmodel.PDDocument;
 import org.springframework.core.io.ByteArrayResource;
 import org.springframework.core.io.Resource;
 import org.springframework.http.HttpHeaders;
+import org.springframework.http.HttpStatus;
 import org.springframework.http.MediaType;
 import org.springframework.http.ResponseEntity;
 import org.springframework.web.bind.annotation.GetMapping;
 import org.springframework.web.bind.annotation.ModelAttribute;
+import org.springframework.web.bind.annotation.PostMapping;
+import org.springframework.web.bind.annotation.RequestBody;
 import org.springframework.web.bind.annotation.RequestMapping;
+import org.springframework.web.bind.annotation.RequestParam;
 import org.springframework.web.bind.annotation.RestController;
 import org.springframework.web.multipart.MultipartFile;
+import org.springframework.web.server.ResponseStatusException;
 
 import io.swagger.v3.oas.annotations.Operation;
 import io.swagger.v3.oas.annotations.tags.Tag;
@@ -29,16 +39,35 @@ import lombok.extern.slf4j.Slf4j;
 
 import stirling.software.common.annotations.AutoJobPostMapping;
 import stirling.software.common.enumeration.ResourceWeight;
+import stirling.software.common.service.CustomPDFDocumentFactory;
+import stirling.software.common.util.FormUtils;
 import stirling.software.common.util.GeneralUtils;
+import stirling.software.common.util.TempFile;
+import stirling.software.common.util.TempFileManager;
 import stirling.software.common.util.WebResponseUtils;
+import stirling.software.proprietary.model.api.docparse.ChunkDocumentApiRequest;
+import stirling.software.proprietary.model.api.docparse.ExtractFieldsApiRequest;
 import stirling.software.proprietary.model.api.docparse.ExtractTablesApiRequest;
+import stirling.software.proprietary.model.api.docparse.ParseDocumentApiRequest;
+import stirling.software.proprietary.model.api.docparse.RagAskApiRequest;
 import stirling.software.proprietary.model.api.docparse.RagIngestApiRequest;
+import stirling.software.proprietary.model.api.docparse.RagSearchApiRequest;
+import stirling.software.proprietary.model.api.docparse.SmartSplitApiRequest;
+import stirling.software.proprietary.model.api.docparse.SuggestSchemaApiRequest;
+import stirling.software.proprietary.model.docparse.ChunkDocumentResponse;
 import stirling.software.proprietary.model.docparse.DocChunk;
 import stirling.software.proprietary.model.docparse.DocTable;
 import stirling.software.proprietary.model.docparse.DocparseCapabilitiesView;
 import stirling.software.proprietary.model.docparse.DocparseMode;
+import stirling.software.proprietary.model.docparse.ExtractFieldsResponse;
 import stirling.software.proprietary.model.docparse.ExtractTablesResponse;
+import stirling.software.proprietary.model.docparse.FillDocxResponse;
+import stirling.software.proprietary.model.docparse.ParseDocumentResponse;
 import stirling.software.proprietary.model.docparse.RagIngestResponse;
+import stirling.software.proprietary.model.docparse.RagStatsView;
+import stirling.software.proprietary.model.docparse.SmartSplitResponse;
+import stirling.software.proprietary.model.docparse.SplitPart;
+import stirling.software.proprietary.model.docparse.SuggestSchemaResponse;
 import stirling.software.proprietary.service.AiToolResponseHeaders;
 import stirling.software.proprietary.service.DocParseService;
 
@@ -63,7 +92,14 @@ public class DocParseController {
 
     private static final MediaType CSV = MediaType.parseMediaType("text/csv");
 
+    private static final MediaType MARKDOWN = MediaType.parseMediaType("text/markdown");
+    private static final MediaType DOCX =
+            MediaType.parseMediaType(
+                    "application/vnd.openxmlformats-officedocument.wordprocessingml.document");
+
     private final DocParseService docParseService;
+    private final CustomPDFDocumentFactory pdfDocumentFactory;
+    private final TempFileManager tempFileManager;
     private final ObjectMapper objectMapper;
 
     @AutoJobPostMapping(
@@ -122,6 +158,189 @@ public class DocParseController {
         return ResponseEntity.ok().headers(headers).body(new ByteArrayResource(zip));
     }
 
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/extract-fields",
+            resourceWeight = ResourceWeight.LARGE_WEIGHT)
+    @Operation(
+            summary = "Extract typed fields from a document (pipeline shape)",
+            description =
+                    "Extracts the fields described by the JSON Schema and returns the ORIGINAL PDF"
+                            + " unchanged as the body, with the extraction JSON in the"
+                            + " X-Stirling-Tool-Report header so policy pipelines pick it up as the"
+                            + " step report. Use /extract-fields/json for the raw JSON."
+                            + " Input:PDF Output:PDF Type:SISO")
+    public ResponseEntity<Resource> extractFields(@ModelAttribute ExtractFieldsApiRequest request)
+            throws IOException {
+        MultipartFile file = request.getFileInput();
+        ExtractFieldsResponse result =
+                docParseService.extractFields(
+                        file,
+                        request.getFieldsSchema(),
+                        DocparseMode.fromWire(request.getMode()),
+                        request.getInstructions());
+        byte[] original = file.getBytes();
+        HttpHeaders headers = new HttpHeaders();
+        headers.setContentType(MediaType.APPLICATION_PDF);
+        headers.setContentDispositionFormData("attachment", DocParseService.fileName(file));
+        headers.setContentLength(original.length);
+        headers.set(AiToolResponseHeaders.TOOL_REPORT, objectMapper.writeValueAsString(result));
+        return ResponseEntity.ok().headers(headers).body(new ByteArrayResource(original));
+    }
+
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/extract-fields/json",
+            resourceWeight = ResourceWeight.LARGE_WEIGHT)
+    @Operation(
+            summary = "Extract typed fields from a document (JSON)",
+            description =
+                    "Extracts the fields described by the JSON Schema and returns the extraction"
+                            + " result (fields, confidence, citations) as JSON."
+                            + " Input:PDF Output:JSON Type:SISO")
+    public ResponseEntity<ExtractFieldsResponse> extractFieldsJson(
+            @ModelAttribute ExtractFieldsApiRequest request) throws IOException {
+        return ResponseEntity.ok(
+                docParseService.extractFields(
+                        request.getFileInput(),
+                        request.getFieldsSchema(),
+                        DocparseMode.fromWire(request.getMode()),
+                        request.getInstructions()));
+    }
+
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/suggest-schema",
+            resourceWeight = ResourceWeight.LARGE_WEIGHT)
+    @Operation(
+            summary = "Suggest an extraction schema for a document",
+            description =
+                    "Reads the document and proposes the fields worth extracting (name, type,"
+                            + " description), ready to feed into /extract-fields as a JSON Schema."
+                            + " Input:PDF Output:JSON Type:SISO")
+    public ResponseEntity<SuggestSchemaResponse> suggestSchema(
+            @ModelAttribute SuggestSchemaApiRequest request) throws IOException {
+        return ResponseEntity.ok(
+                docParseService.suggestSchema(request.getFileInput(), request.getMaxFields()));
+    }
+
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/parse-document",
+            resourceWeight = ResourceWeight.XLARGE_WEIGHT)
+    @Operation(
+            summary = "Parse a document into structured blocks, tables, and markdown",
+            description =
+                    "Parses the PDF into layout blocks, tables, and a markdown rendering. The"
+                            + " basic tier reads the text layer; the advanced tier (docparse addon)"
+                            + " adds OCR, real table structure, and bounding boxes."
+                            + " Input:PDF Output:JSON Type:SISO")
+    public ResponseEntity<?> parseDocument(@ModelAttribute ParseDocumentApiRequest request)
+            throws IOException {
+        ParseDocumentResponse result =
+                docParseService.parse(
+                        request.getFileInput(),
+                        DocparseMode.fromWire(request.getMode()),
+                        request.isWithOcr());
+        if ("markdown".equalsIgnoreCase(request.getOutputFormat())) {
+            return WebResponseUtils.bytesToWebResponse(
+                    result.markdown().getBytes(StandardCharsets.UTF_8),
+                    outputName(request.getFileInput(), "_parsed.md"),
+                    MARKDOWN);
+        }
+        return ResponseEntity.ok(result);
+    }
+
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/smart-split",
+            resourceWeight = ResourceWeight.LARGE_WEIGHT)
+    @Operation(
+            summary = "Split a document at content-derived boundaries",
+            description =
+                    "Asks the engine where sub-documents start (per the natural-language rule) and"
+                            + " returns a ZIP with one PDF per part, named from the part labels."
+                            + " Input:PDF Output:ZIP-PDF Type:SIMO")
+    public ResponseEntity<Resource> smartSplit(@ModelAttribute SmartSplitApiRequest request)
+            throws IOException {
+        MultipartFile file = request.getFileInput();
+        SmartSplitResponse split =
+                docParseService.split(file, request.getRule(), request.getMaxParts());
+        if (split.parts().isEmpty()) {
+            throw new ResponseStatusException(
+                    HttpStatus.UNPROCESSABLE_ENTITY,
+                    "The split rule produced no parts for this document");
+        }
+        TempFile zipTempFile = tempFileManager.createManagedTempFile(".zip");
+        try {
+            try (TempFile sourceTempFile = new TempFile(tempFileManager, ".pdf")) {
+                Files.copy(
+                        file.getInputStream(),
+                        sourceTempFile.getPath(),
+                        StandardCopyOption.REPLACE_EXISTING);
+                try (ZipOutputStream zipOut =
+                        new ZipOutputStream(Files.newOutputStream(zipTempFile.getPath()))) {
+                    writeParts(sourceTempFile, split.parts(), zipOut);
+                }
+            }
+            return WebResponseUtils.zipFileToWebResponse(
+                    zipTempFile,
+                    GeneralUtils.generateFilename(file.getOriginalFilename(), "_split.zip"));
+        } catch (Exception e) {
+            zipTempFile.close();
+            throw e;
+        }
+    }
+
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/chunk-document",
+            resourceWeight = ResourceWeight.MEDIUM_WEIGHT)
+    @Operation(
+            summary = "Chunk a document for RAG",
+            description =
+                    "Splits the document text into overlapping chunks with page spans and (advanced"
+                            + " tier) heading breadcrumbs. Input:PDF Output:JSON Type:SISO")
+    public ResponseEntity<ChunkDocumentResponse> chunkDocument(
+            @ModelAttribute ChunkDocumentApiRequest request) throws IOException {
+        return ResponseEntity.ok(
+                docParseService.chunk(
+                        request.getFileInput(),
+                        request.getChunkSize(),
+                        request.getOverlap(),
+                        DocparseMode.fromWire(request.getMode())));
+    }
+
+    @AutoJobPostMapping(
+            consumes = MediaType.MULTIPART_FORM_DATA_VALUE,
+            value = "/fill-template",
+            resourceWeight = ResourceWeight.SMALL_WEIGHT)
+    @Operation(
+            summary = "Fill a DOCX template with JSON data",
+            description =
+                    "Replaces the template's placeholders with values from the JSON object and"
+                            + " returns the filled DOCX. Replacement counts and missing keys ride"
+                            + " the X-Stirling-Tool-Report header."
+                            + " Input:DOCX Output:DOCX Type:SISO")
+    public ResponseEntity<Resource> fillTemplate(
+            @RequestParam("templateFile") MultipartFile templateFile,
+            @RequestParam("data") String data)
+            throws IOException {
+        FillDocxResponse result = docParseService.fillDocx(templateFile, data);
+        byte[] filled = Base64.getDecoder().decode(result.docxBase64());
+        HttpHeaders headers = new HttpHeaders();
+        headers.setContentType(DOCX);
+        headers.setContentDispositionFormData(
+                "attachment",
+                GeneralUtils.generateFilename(templateFile.getOriginalFilename(), "_filled.docx"));
+        headers.setContentLength(filled.length);
+        headers.set(
+                AiToolResponseHeaders.TOOL_REPORT,
+                objectMapper.writeValueAsString(
+                        new FillDocxResponse("", result.replaced(), result.missing())));
+        return ResponseEntity.ok().headers(headers).body(new ByteArrayResource(filled));
+    }
+
     @GetMapping("/capabilities")
     @Operation(
             summary = "DocParse capability summary",
@@ -154,6 +373,52 @@ public class DocParseController {
                 CSV);
     }
 
+    @GetMapping("/rag-stats")
+    @Operation(
+            summary = "RAG store statistics",
+            description =
+                    "The engine's document-store totals (backend, documents, chunks, embedding"
+                            + " model) merged with the DocParse capability fields. Answers with"
+                            + " zeros and engineReachable=false when the engine is down.")
+    public ResponseEntity<RagStatsView> ragStats() {
+        return ResponseEntity.ok(docParseService.ragStats());
+    }
+
+    @GetMapping("/rag-documents")
+    @Operation(
+            summary = "List documents in the RAG store",
+            description =
+                    "Engine passthrough of the caller-visible indexed documents (documentId,"
+                            + " source, chunk count).")
+    public ResponseEntity<String> ragDocuments() throws IOException {
+        return jsonPassthrough(docParseService.ragDocuments());
+    }
+
+    @PostMapping(value = "/rag-search", consumes = MediaType.APPLICATION_JSON_VALUE)
+    @Operation(
+            summary = "Semantic search over the RAG store",
+            description =
+                    "Searches the caller-visible indexed documents and returns the top passages"
+                            + " with scores, page spans, and heading breadcrumbs.")
+    public ResponseEntity<String> ragSearch(@RequestBody RagSearchApiRequest request)
+            throws IOException {
+        return jsonPassthrough(docParseService.ragSearch(request.getQuery(), request.getTopK()));
+    }
+
+    @PostMapping(value = "/rag-ask", consumes = MediaType.APPLICATION_JSON_VALUE)
+    @Operation(
+            summary = "Ask a question over the RAG store",
+            description =
+                    "Answers the question from the caller-visible indexed documents and returns"
+                            + " the answer with its supporting passages.")
+    public ResponseEntity<String> ragAsk(@RequestBody RagAskApiRequest request) throws IOException {
+        return jsonPassthrough(docParseService.ragAsk(request.getQuestion(), request.getTopK()));
+    }
+
+    private static ResponseEntity<String> jsonPassthrough(String engineJson) {
+        return ResponseEntity.ok().contentType(MediaType.APPLICATION_JSON).body(engineJson);
+    }
+
     /** Original + requested corpus files in one ZIP, so destinations receive them together. */
     private byte[] exportZip(
             String fileName, byte[] original, RagIngestResponse result, RagIngestApiRequest request)
@@ -209,6 +474,40 @@ public class DocParseController {
         return dot > 0 ? fileName.substring(0, dot) : fileName;
     }
 
+    private void writeParts(TempFile sourceTempFile, List<SplitPart> parts, ZipOutputStream zipOut)
+            throws IOException {
+        for (int i = 0; i < parts.size(); i++) {
+            SplitPart part = parts.get(i);
+            // Load per part and remove pages outside the range: avoids the PDFBox cross-document
+            // addPage pitfalls while keeping shared resources intact.
+            try (PDDocument partDoc = pdfDocumentFactory.load(sourceTempFile.getFile())) {
+                int pageCount = partDoc.getNumberOfPages();
+                int start = Math.clamp(part.startPage(), 1, pageCount);
+                int end = Math.clamp(part.endPage(), start, pageCount);
+                for (int p = pageCount - 1; p >= 0; p--) {
+                    int pageNumber = p + 1;
+                    if (pageNumber < start || pageNumber > end) {
+                        partDoc.removePage(p);
+                    }
+                }
+                FormUtils.pruneOrphanedFormFields(partDoc);
+                zipOut.putNextEntry(new ZipEntry(partEntryName(i, part)));
+                partDoc.save(zipOut);
+                zipOut.closeEntry();
+            }
+        }
+    }
+
+    private static String partEntryName(int index, SplitPart part) {
+        String label = part.label() == null ? "" : part.label().trim();
+        String sanitized = label.replaceAll("[^A-Za-z0-9 ._-]", "_").replaceAll("\\s+", "_");
+        if (sanitized.isBlank() || sanitized.chars().allMatch(c -> c == '_' || c == '.')) {
+            sanitized = "part";
+        }
+        // Index prefix keeps entries unique even when labels repeat.
+        return String.format(Locale.ROOT, "%02d_%s.pdf", index + 1, sanitized);
+    }
+
     private static String tablesToCsv(List<DocTable> tables) throws IOException {
         CSVFormat format = CSVFormat.EXCEL.builder().setEscape('"').build();
         StringWriter writer = new StringWriter();

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/ChunkDocumentApiRequest.java

@@ -0,0 +1,27 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import stirling.software.common.model.api.PDFFile;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+public class ChunkDocumentApiRequest extends PDFFile {
+
+    @Schema(description = "Target chunk size in characters (64-32768)", defaultValue = "512")
+    private int chunkSize = 512;
+
+    @Schema(
+            description = "Overlap between adjacent chunks in characters (0-4096)",
+            defaultValue = "64")
+    private int overlap = 64;
+
+    @Schema(
+            description = "Tier to use: 'auto' picks per document, or force 'basic'/'advanced'",
+            allowableValues = {"auto", "basic", "advanced"},
+            defaultValue = "auto")
+    private String mode = "auto";
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/ExtractFieldsApiRequest.java

@@ -0,0 +1,29 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import stirling.software.common.model.api.PDFFile;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+public class ExtractFieldsApiRequest extends PDFFile {
+
+    @Schema(
+            description = "JSON Schema object describing the fields to extract, as a JSON string",
+            requiredMode = Schema.RequiredMode.REQUIRED,
+            example =
+                    "{\"type\":\"object\",\"properties\":{\"invoiceNumber\":{\"type\":\"string\"}}}")
+    private String fieldsSchema;
+
+    @Schema(
+            description = "Tier to use: 'auto' picks per document, or force 'basic'/'advanced'",
+            allowableValues = {"auto", "basic", "advanced"},
+            defaultValue = "auto")
+    private String mode = "auto";
+
+    @Schema(description = "Optional natural-language guidance for the extraction")
+    private String instructions;
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/ParseDocumentApiRequest.java

@@ -0,0 +1,30 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import stirling.software.common.model.api.PDFFile;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+public class ParseDocumentApiRequest extends PDFFile {
+
+    @Schema(
+            description = "Tier to use: 'auto' picks per document, or force 'basic'/'advanced'",
+            allowableValues = {"auto", "basic", "advanced"},
+            defaultValue = "auto")
+    private String mode = "auto";
+
+    @Schema(
+            description = "Apply OCR when parsing scanned pages (advanced tier only)",
+            defaultValue = "true")
+    private boolean withOcr = true;
+
+    @Schema(
+            description = "Response format: full JSON result or the markdown rendering only",
+            allowableValues = {"json", "markdown"},
+            defaultValue = "json")
+    private String outputFormat = "json";
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/RagAskApiRequest.java

@@ -0,0 +1,17 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+
+@Data
+public class RagAskApiRequest {
+
+    @Schema(
+            description = "Question to answer from the indexed documents",
+            requiredMode = Schema.RequiredMode.REQUIRED)
+    private String question;
+
+    @Schema(description = "Number of passages to ground the answer on (1-20)", defaultValue = "5")
+    private int topK = 5;
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/RagSearchApiRequest.java

@@ -0,0 +1,17 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+
+@Data
+public class RagSearchApiRequest {
+
+    @Schema(
+            description = "Natural-language search query",
+            requiredMode = Schema.RequiredMode.REQUIRED)
+    private String query;
+
+    @Schema(description = "Number of passages to return (1-50)", defaultValue = "10")
+    private int topK = 10;
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/SmartSplitApiRequest.java

@@ -0,0 +1,21 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import stirling.software.common.model.api.PDFFile;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+public class SmartSplitApiRequest extends PDFFile {
+
+    @Schema(
+            description = "Natural-language boundary rule, e.g. 'split where a new invoice starts'",
+            requiredMode = Schema.RequiredMode.REQUIRED)
+    private String rule;
+
+    @Schema(description = "Maximum number of parts to produce (1-500)", defaultValue = "50")
+    private int maxParts = 50;
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/api/docparse/SuggestSchemaApiRequest.java

@@ -0,0 +1,16 @@
+package stirling.software.proprietary.model.api.docparse;
+
+import io.swagger.v3.oas.annotations.media.Schema;
+
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import stirling.software.common.model.api.PDFFile;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+public class SuggestSchemaApiRequest extends PDFFile {
+
+    @Schema(description = "Maximum number of fields to suggest (1-20)", defaultValue = "10")
+    private int maxFields = 10;
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ChunkDocumentRequest.java

@@ -0,0 +1,14 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+import stirling.software.proprietary.model.api.ai.AiPageText;
+
+/** Engine request for {@code POST /api/v1/docparse/chunk}. */
+public record ChunkDocumentRequest(
+        String fileName,
+        List<AiPageText> pages,
+        String contentBase64,
+        int chunkSize,
+        int overlap,
+        DocparseMode mode) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ChunkDocumentResponse.java

@@ -0,0 +1,11 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/** Engine response for {@code POST /api/v1/docparse/chunk}. */
+public record ChunkDocumentResponse(DocparseTier mode, List<DocChunk> chunks) {
+
+    public ChunkDocumentResponse {
+        chunks = chunks == null ? List.of() : chunks;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/DocBlock.java

@@ -0,0 +1,9 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/**
+ * One layout block. {@code bbox} is [x0, y0, x1, y1] normalized to 0..1 with a top-left origin;
+ * {@code null} in basic tier (no layout model ran). Mirrors {@code docparse.py DocBlock}.
+ */
+public record DocBlock(String type, String text, int page, List<Double> bbox, Double confidence) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/DocumentStoreStats.java

@@ -0,0 +1,5 @@
+package stirling.software.proprietary.model.docparse;
+
+/** Engine response for {@code GET /api/v1/documents/stats}: the RAG document store totals. */
+public record DocumentStoreStats(
+        String backend, long documents, long chunks, String embeddingModel) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ExtractFieldsRequest.java

@@ -0,0 +1,19 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+import stirling.software.proprietary.model.api.ai.AiPageText;
+
+import tools.jackson.databind.JsonNode;
+
+/**
+ * Engine request for {@code POST /api/v1/docparse/extract}. {@code pages} drives the basic tier
+ * (Java-extracted text); {@code contentBase64} lets the advanced tier parse the raw file itself.
+ */
+public record ExtractFieldsRequest(
+        String fileName,
+        JsonNode fieldsSchema,
+        List<AiPageText> pages,
+        String contentBase64,
+        DocparseMode mode,
+        String instructions) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ExtractFieldsResponse.java

@@ -0,0 +1,12 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/** Engine response for {@code POST /api/v1/docparse/extract}. */
+public record ExtractFieldsResponse(
+        DocparseTier mode, List<ExtractedField> fields, double overallConfidence) {
+
+    public ExtractFieldsResponse {
+        fields = fields == null ? List.of() : fields;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ExtractedField.java

@@ -0,0 +1,16 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+import tools.jackson.databind.JsonNode;
+
+/**
+ * One extracted field with confidence and citations. Mirrors {@code docparse.py ExtractedField}.
+ */
+public record ExtractedField(
+        String name, JsonNode value, double confidence, List<FieldCitation> citations) {
+
+    public ExtractedField {
+        citations = citations == null ? List.of() : citations;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/FieldCitation.java

@@ -0,0 +1,11 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/**
+ * Where a value came from. {@code quote} is always set; {@code bbox} only when a layout parse ran
+ * (advanced tier); offsets index into the cited page's text. Mirrors {@code docparse.py
+ * FieldCitation}.
+ */
+public record FieldCitation(
+        Integer page, List<Double> bbox, String quote, Integer startOffset, Integer endOffset) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/FillDocxRequest.java

@@ -0,0 +1,6 @@
+package stirling.software.proprietary.model.docparse;
+
+import tools.jackson.databind.JsonNode;
+
+/** Engine request for {@code POST /api/v1/docparse/fill-docx}. */
+public record FillDocxRequest(String templateBase64, JsonNode data) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/FillDocxResponse.java

@@ -0,0 +1,11 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/** Engine response for {@code POST /api/v1/docparse/fill-docx}. */
+public record FillDocxResponse(String docxBase64, int replaced, List<String> missing) {
+
+    public FillDocxResponse {
+        missing = missing == null ? List.of() : missing;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ParseDocumentRequest.java

@@ -0,0 +1,4 @@
+package stirling.software.proprietary.model.docparse;
+
+/** Engine request for {@code POST /api/v1/docparse/parse}. */
+public record ParseDocumentRequest(String fileName, String contentBase64, boolean withOcr) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/ParseDocumentResponse.java

@@ -0,0 +1,21 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/**
+ * Engine response for {@code POST /api/v1/docparse/parse}; also produced by the Java basic tier.
+ */
+public record ParseDocumentResponse(
+        DocparseTier mode,
+        int pages,
+        List<DocBlock> blocks,
+        List<DocTable> tables,
+        String markdown,
+        boolean ocrApplied) {
+
+    public ParseDocumentResponse {
+        blocks = blocks == null ? List.of() : blocks;
+        tables = tables == null ? List.of() : tables;
+        markdown = markdown == null ? "" : markdown;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/RagAskRequest.java

@@ -0,0 +1,4 @@
+package stirling.software.proprietary.model.docparse;
+
+/** Engine request for {@code POST /api/v1/documents/ask}: grounded Q&A over the RAG store. */
+public record RagAskRequest(String question, int topK) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/RagSearchRequest.java

@@ -0,0 +1,4 @@
+package stirling.software.proprietary.model.docparse;
+
+/** Engine request for {@code POST /api/v1/documents/search}: semantic search over the RAG store. */
+public record RagSearchRequest(String query, int topK) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/RagStatsView.java

@@ -0,0 +1,15 @@
+package stirling.software.proprietary.model.docparse;
+
+/**
+ * Merged RAG store view served by {@code GET /api/v1/docparse/rag-stats} (Java side): the engine's
+ * document-store totals plus the cached DocParse capability fields. When the engine is unreachable
+ * the totals are zero and {@code engineReachable} is false.
+ */
+public record RagStatsView(
+        String backend,
+        long documents,
+        long chunks,
+        String embeddingModel,
+        boolean advancedInstalled,
+        String doclingVersion,
+        boolean engineReachable) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/SmartSplitRequest.java

@@ -0,0 +1,9 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+import stirling.software.proprietary.model.api.ai.AiPageText;
+
+/** Engine request for {@code POST /api/v1/docparse/split}. */
+public record SmartSplitRequest(
+        String fileName, String rule, List<AiPageText> pages, int maxParts) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/SmartSplitResponse.java

@@ -0,0 +1,11 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/** Engine response for {@code POST /api/v1/docparse/split}. */
+public record SmartSplitResponse(List<SplitPart> parts) {
+
+    public SmartSplitResponse {
+        parts = parts == null ? List.of() : parts;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/SplitPart.java

@@ -0,0 +1,4 @@
+package stirling.software.proprietary.model.docparse;
+
+/** One sub-document page range (1-based, inclusive). Mirrors {@code docparse.py SplitPart}. */
+public record SplitPart(int startPage, int endPage, String label, double confidence) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/SuggestSchemaRequest.java

@@ -0,0 +1,12 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+import stirling.software.proprietary.model.api.ai.AiPageText;
+
+/**
+ * Engine request for {@code POST /api/v1/docparse/suggest-schema}. {@code pages} drives the basic
+ * tier (Java-extracted text); {@code contentBase64} lets the advanced tier parse the raw file.
+ */
+public record SuggestSchemaRequest(
+        String fileName, List<AiPageText> pages, String contentBase64, int maxFields) {}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/SuggestSchemaResponse.java

@@ -0,0 +1,11 @@
+package stirling.software.proprietary.model.docparse;
+
+import java.util.List;
+
+/** Engine response for {@code POST /api/v1/docparse/suggest-schema}. */
+public record SuggestSchemaResponse(DocparseTier mode, List<SuggestedField> fields) {
+
+    public SuggestSchemaResponse {
+        fields = fields == null ? List.of() : fields;
+    }
+}

app/proprietary/src/main/java/stirling/software/proprietary/model/docparse/SuggestedField.java

@@ -0,0 +1,4 @@
+package stirling.software.proprietary.model.docparse;
+
+/** One field the engine proposes for an extraction schema. */
+public record SuggestedField(String name, String type, String description) {}

app/proprietary/src/main/java/stirling/software/proprietary/service/DocParseService.java

@@ -20,15 +20,33 @@ import stirling.software.common.model.ApplicationProperties;
 import stirling.software.common.service.CustomPDFDocumentFactory;
 import stirling.software.common.service.UserServiceInterface;
 import stirling.software.proprietary.model.api.ai.AiPageText;
+import stirling.software.proprietary.model.docparse.ChunkDocumentRequest;
+import stirling.software.proprietary.model.docparse.ChunkDocumentResponse;
+import stirling.software.proprietary.model.docparse.DocBlock;
 import stirling.software.proprietary.model.docparse.DocparseCapabilities;
 import stirling.software.proprietary.model.docparse.DocparseCapabilitiesView;
 import stirling.software.proprietary.model.docparse.DocparseMode;
 import stirling.software.proprietary.model.docparse.DocparseTier;
+import stirling.software.proprietary.model.docparse.DocumentStoreStats;
+import stirling.software.proprietary.model.docparse.ExtractFieldsRequest;
+import stirling.software.proprietary.model.docparse.ExtractFieldsResponse;
 import stirling.software.proprietary.model.docparse.ExtractTablesRequest;
 import stirling.software.proprietary.model.docparse.ExtractTablesResponse;
+import stirling.software.proprietary.model.docparse.FillDocxRequest;
+import stirling.software.proprietary.model.docparse.FillDocxResponse;
+import stirling.software.proprietary.model.docparse.ParseDocumentRequest;
+import stirling.software.proprietary.model.docparse.ParseDocumentResponse;
+import stirling.software.proprietary.model.docparse.RagAskRequest;
 import stirling.software.proprietary.model.docparse.RagIngestRequest;
 import stirling.software.proprietary.model.docparse.RagIngestResponse;
+import stirling.software.proprietary.model.docparse.RagSearchRequest;
+import stirling.software.proprietary.model.docparse.RagStatsView;
+import stirling.software.proprietary.model.docparse.SmartSplitRequest;
+import stirling.software.proprietary.model.docparse.SmartSplitResponse;
+import stirling.software.proprietary.model.docparse.SuggestSchemaRequest;
+import stirling.software.proprietary.model.docparse.SuggestSchemaResponse;
 
+import tools.jackson.databind.JsonNode;
 import tools.jackson.databind.ObjectMapper;
 
 /**
@@ -41,8 +59,18 @@ import tools.jackson.databind.ObjectMapper;
 @Service
 public class DocParseService {
 
-    private static final String RAG_INGEST_ENDPOINT = "/api/v1/docparse/rag-ingest";
+    private static final String PARSE_ENDPOINT = "/api/v1/docparse/parse";
+    private static final String EXTRACT_ENDPOINT = "/api/v1/docparse/extract";
+    private static final String SPLIT_ENDPOINT = "/api/v1/docparse/split";
+    private static final String CHUNK_ENDPOINT = "/api/v1/docparse/chunk";
     private static final String TABLES_ENDPOINT = "/api/v1/docparse/tables";
+    private static final String FILL_DOCX_ENDPOINT = "/api/v1/docparse/fill-docx";
+    private static final String SUGGEST_SCHEMA_ENDPOINT = "/api/v1/docparse/suggest-schema";
+    private static final String RAG_INGEST_ENDPOINT = "/api/v1/docparse/rag-ingest";
+    private static final String DOCUMENT_STATS_ENDPOINT = "/api/v1/documents/stats";
+    private static final String DOCUMENT_LIST_ENDPOINT = "/api/v1/documents/list";
+    private static final String DOCUMENT_SEARCH_ENDPOINT = "/api/v1/documents/search";
+    private static final String DOCUMENT_ASK_ENDPOINT = "/api/v1/documents/ask";
 
     /** Below this average of extractable chars per page the document is treated as scanned. */
     static final int SCANNED_AVG_CHARS_PER_PAGE = 100;
@@ -97,6 +125,72 @@ public class DocParseService {
                 capabilities.doclingVersion());
     }
 
+    public ParseDocumentResponse parse(
+            MultipartFile file, DocparseMode requestedMode, boolean withOcr) throws IOException {
+        requireEnabled();
+        DocparseTier tier;
+        try (PDDocument document = pdfDocumentFactory.load(file, true)) {
+            tier =
+                    resolveTier(
+                            requestedMode,
+                            capabilityService.capabilities(),
+                            false,
+                            looksScanned(document));
+            if (tier == DocparseTier.BASIC) {
+                return basicParse(document);
+            }
+        }
+        ParseDocumentRequest request =
+                new ParseDocumentRequest(fileName(file), encodeBase64(file), withOcr);
+        String responseJson =
+                aiEngineClient.postLongRunning(
+                        PARSE_ENDPOINT, objectMapper.writeValueAsString(request), currentUserId());
+        return objectMapper.readValue(responseJson, ParseDocumentResponse.class);
+    }
+
+    public SmartSplitResponse split(MultipartFile file, String rule, int maxParts)
+            throws IOException {
+        requireEnabled();
+        if (rule == null || rule.isBlank()) {
+            throw new ResponseStatusException(HttpStatus.BAD_REQUEST, "A split rule is required");
+        }
+        List<AiPageText> pages;
+        try (PDDocument document = pdfDocumentFactory.load(file, true)) {
+            pages = extractPages(document);
+        }
+        SmartSplitRequest request =
+                new SmartSplitRequest(fileName(file), rule, pages, Math.clamp(maxParts, 1, 500));
+        String responseJson =
+                aiEngineClient.post(
+                        SPLIT_ENDPOINT, objectMapper.writeValueAsString(request), currentUserId());
+        return objectMapper.readValue(responseJson, SmartSplitResponse.class);
+    }
+
+    public ChunkDocumentResponse chunk(
+            MultipartFile file, int chunkSize, int overlap, DocparseMode mode) throws IOException {
+        requireEnabled();
+        List<AiPageText> pages;
+        DocparseTier tier;
+        try (PDDocument document = pdfDocumentFactory.load(file, true)) {
+            pages = extractPages(document);
+            tier =
+                    resolveTier(
+                            mode, capabilityService.capabilities(), false, looksScanned(document));
+        }
+        ChunkDocumentRequest request =
+                new ChunkDocumentRequest(
+                        fileName(file),
+                        pages,
+                        tier == DocparseTier.ADVANCED ? encodeBase64(file) : null,
+                        Math.clamp(chunkSize, 64, 32_768),
+                        Math.clamp(overlap, 0, 4_096),
+                        toMode(tier));
+        String responseJson =
+                aiEngineClient.postLongRunning(
+                        CHUNK_ENDPOINT, objectMapper.writeValueAsString(request), currentUserId());
+        return objectMapper.readValue(responseJson, ChunkDocumentResponse.class);
+    }
+
     /**
      * Chunk, embed, and index the document into the engine's RAG store, and/or echo the parsed
      * content back for corpus export. Text extraction and tier routing happen here; the engine
@@ -182,6 +276,146 @@ public class DocParseService {
         };
     }
 
+    /** Extract the fields described by a JSON Schema, with confidence and citations. */
+    public ExtractFieldsResponse extractFields(
+            MultipartFile file, String fieldsSchemaJson, DocparseMode mode, String instructions)
+            throws IOException {
+        requireEnabled();
+        JsonNode schema = parseJsonObject(fieldsSchemaJson, "fieldsSchema");
+        List<AiPageText> pages;
+        DocparseTier tier;
+        try (PDDocument document = pdfDocumentFactory.load(file, true)) {
+            pages = extractPages(document);
+            tier =
+                    resolveTier(
+                            mode, capabilityService.capabilities(), false, looksScanned(document));
+        }
+        ExtractFieldsRequest request =
+                new ExtractFieldsRequest(
+                        fileName(file),
+                        schema,
+                        pages,
+                        tier == DocparseTier.ADVANCED ? encodeBase64(file) : null,
+                        toMode(tier),
+                        instructions);
+        String responseJson =
+                aiEngineClient.postLongRunning(
+                        EXTRACT_ENDPOINT,
+                        objectMapper.writeValueAsString(request),
+                        currentUserId());
+        return objectMapper.readValue(responseJson, ExtractFieldsResponse.class);
+    }
+
+    /** Propose an extraction schema from the document's first pages. */
+    public SuggestSchemaResponse suggestSchema(MultipartFile file, int maxFields)
+            throws IOException {
+        requireEnabled();
+        List<AiPageText> pages;
+        DocparseTier tier;
+        try (PDDocument document = pdfDocumentFactory.load(file, true)) {
+            pages = extractPages(document);
+            tier =
+                    resolveTier(
+                            DocparseMode.AUTO,
+                            capabilityService.capabilities(),
+                            false,
+                            looksScanned(document));
+        }
+        SuggestSchemaRequest request =
+                new SuggestSchemaRequest(
+                        fileName(file),
+                        pages,
+                        tier == DocparseTier.ADVANCED ? encodeBase64(file) : null,
+                        Math.clamp(maxFields, 1, 20));
+        String responseJson =
+                aiEngineClient.postLongRunning(
+                        SUGGEST_SCHEMA_ENDPOINT,
+                        objectMapper.writeValueAsString(request),
+                        currentUserId());
+        return objectMapper.readValue(responseJson, SuggestSchemaResponse.class);
+    }
+
+    private JsonNode parseJsonObject(String json, String fieldName) {
+        if (json == null || json.isBlank()) {
+            throw new ResponseStatusException(
+                    HttpStatus.BAD_REQUEST, "'" + fieldName + "' is required");
+        }
+        JsonNode node;
+        try {
+            node = objectMapper.readTree(json);
+        } catch (Exception e) {
+            throw new ResponseStatusException(
+                    HttpStatus.BAD_REQUEST, "'" + fieldName + "' is not valid JSON");
+        }
+        if (!node.isObject()) {
+            throw new ResponseStatusException(
+                    HttpStatus.BAD_REQUEST, "'" + fieldName + "' must be a JSON object");
+        }
+        return node;
+    }
+
+    private static void requireNonBlank(String value, String fieldName) {
+        if (value == null || value.isBlank()) {
+            throw new ResponseStatusException(
+                    HttpStatus.BAD_REQUEST, "'" + fieldName + "' is required");
+        }
+    }
+
+    /** Engine RAG store totals merged with the cached capability fields; graceful when down. */
+    public RagStatsView ragStats() {
+        DocparseCapabilities capabilities = capabilityService.capabilities();
+        try {
+            // The engine's documents routes are user-gated; an id-less probe 401s
+            // and would read as "engine offline" in the UI.
+            String json = aiEngineClient.get(DOCUMENT_STATS_ENDPOINT, currentUserId());
+            DocumentStoreStats stats = objectMapper.readValue(json, DocumentStoreStats.class);
+            return new RagStatsView(
+                    stats.backend(),
+                    stats.documents(),
+                    stats.chunks(),
+                    stats.embeddingModel(),
+                    capabilities.advancedInstalled(),
+                    capabilities.doclingVersion(),
+                    true);
+        } catch (Exception e) {
+            log.debug("RAG stats probe failed: {}", e.getMessage());
+            return new RagStatsView(
+                    null,
+                    0,
+                    0,
+                    null,
+                    capabilities.advancedInstalled(),
+                    capabilities.doclingVersion(),
+                    false);
+        }
+    }
+
+    /** Engine document-list passthrough; X-User-Id scopes it to the caller's ACLs. */
+    public String ragDocuments() throws IOException {
+        requireEnabled();
+        return aiEngineClient.get(DOCUMENT_LIST_ENDPOINT, currentUserId());
+    }
+
+    /** Semantic-search passthrough over the caller-visible RAG documents. */
+    public String ragSearch(String query, int topK) throws IOException {
+        requireEnabled();
+        requireNonBlank(query, "query");
+        RagSearchRequest request = new RagSearchRequest(query, Math.clamp(topK, 1, 50));
+        return aiEngineClient.post(
+                DOCUMENT_SEARCH_ENDPOINT,
+                objectMapper.writeValueAsString(request),
+                currentUserId());
+    }
+
+    /** Grounded-answer passthrough; long-running because local models answer slowly. */
+    public String ragAsk(String question, int topK) throws IOException {
+        requireEnabled();
+        requireNonBlank(question, "question");
+        RagAskRequest request = new RagAskRequest(question, Math.clamp(topK, 1, 20));
+        return aiEngineClient.postLongRunning(
+                DOCUMENT_ASK_ENDPOINT, objectMapper.writeValueAsString(request), currentUserId());
+    }
+
     /**
      * The settings mode wins when stricter: a settings {@code basic} always forces basic, a
      * settings {@code advanced} upgrades everything except an explicit basic request.
@@ -242,6 +476,41 @@ public class DocParseService {
         return objectMapper.readValue(responseJson, ExtractTablesResponse.class);
     }
 
+    public FillDocxResponse fillDocx(MultipartFile templateFile, String dataJson)
+            throws IOException {
+        requireEnabled();
+        JsonNode data = parseJsonObject(dataJson, "data");
+        FillDocxRequest request =
+                new FillDocxRequest(
+                        Base64.getEncoder().encodeToString(templateFile.getBytes()), data);
+        String responseJson =
+                aiEngineClient.post(
+                        FILL_DOCX_ENDPOINT,
+                        objectMapper.writeValueAsString(request),
+                        currentUserId());
+        return objectMapper.readValue(responseJson, FillDocxResponse.class);
+    }
+
+    /** Basic tier parse: PDFBox text layer only, one paragraph block per non-blank page. */
+    ParseDocumentResponse basicParse(PDDocument document) throws IOException {
+        int pageCount = document.getNumberOfPages();
+        List<DocBlock> blocks = new ArrayList<>();
+        StringBuilder markdown = new StringBuilder();
+        for (int page = 1; page <= pageCount; page++) {
+            String text = pdfContentExtractor.extractPageTextRaw(document, page);
+            if (text == null || text.isBlank()) {
+                continue;
+            }
+            blocks.add(new DocBlock("paragraph", text, page, null, null));
+            if (!markdown.isEmpty()) {
+                markdown.append("\n\n");
+            }
+            markdown.append(text);
+        }
+        return new ParseDocumentResponse(
+                DocparseTier.BASIC, pageCount, blocks, List.of(), markdown.toString(), false);
+    }
+
     /** Extract per-page text for the engine, capped by the shared aiEngine limits. */
     List<AiPageText> extractPages(PDDocument document) throws IOException {
         ApplicationProperties.AiEngine.Limits limits =

engine/pyproject.toml

@@ -14,6 +14,8 @@ dependencies = [
     "pydantic-ai-slim[voyageai]>=1.99.0,<2.0.0",
     "pydantic-settings>=2.0.0",
     "python-dotenv>=1.2.1",
+    # Small (MIT) and always installed: DOCX template filling needs no addon.
+    "python-docx>=1.1.2",
     "sqlite-vec>=0.1.6",
     "uvicorn>=0.35.0",
     "opentelemetry-sdk>=1.39.0",

engine/src/stirling/agents/__init__.py

@@ -2,6 +2,7 @@
 
 from .document_classifier import DocumentClassifierAgent
 from .execution import ExecutionPlanningAgent
+from .knowledge_ask import KnowledgeAskAgent
 from .orchestrator import OrchestratorAgent
 from .pdf_create import PdfCreateAgent
 from .pdf_edit import PdfEditAgent, PdfEditParameterSelector, PdfEditPlanSelection
@@ -12,6 +13,7 @@ from .user_spec import UserSpecAgent
 __all__ = [
     "DocumentClassifierAgent",
     "ExecutionPlanningAgent",
+    "KnowledgeAskAgent",
     "OrchestratorAgent",
     "PdfCreateAgent",
     "PdfEditAgent",

engine/src/stirling/agents/knowledge_ask.py

@@ -0,0 +1,134 @@
+"""Grounded Q&A over the caller's stored documents.
+
+Retrieval runs the same cross-collection search as ``POST /documents/search``;
+one smart-model pass then answers only from the retrieved passages, citing
+document and page inline. No retrieval hit means a plain "not found" answer.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from pydantic import Field
+from pydantic_ai import Agent
+
+from stirling.agents.output_mode import output_retries, structured_output
+from stirling.contracts import AskDocumentsRequest, AskDocumentsResponse, DocumentPassage
+from stirling.documents import CollectionSearchHit
+from stirling.documents.service import PAGE_NUMBER_METADATA_KEY
+from stirling.models import ApiModel, PrincipalId
+from stirling.services import AppRuntime
+
+logger = logging.getLogger(__name__)
+
+# Metadata keys written by docparse rag-ingest (_chunk_metadata) for structure-aware chunks.
+_PAGE_START_KEY = "page_start"
+_PAGE_END_KEY = "page_end"
+_HEADING_PATH_KEY = "heading_path"
+_HEADING_PATH_SEPARATOR = " > "
+
+_SYSTEM_PROMPT = (
+    "You answer questions using ONLY the numbered passages you are given.\n"
+    "\n"
+    "Rules:\n"
+    "- Every statement must come from the passages. Never use outside knowledge, never guess.\n"
+    "- Cite the document and page inline right after each fact, "
+    'e.g. "(invoice.pdf p.2)" or "(report.pdf p.4-6)", using the names and pages '
+    "shown in each passage header.\n"
+    "- If the passages do not answer the question, say plainly that the stored "
+    "documents do not cover it. Do not attempt a partial guess.\n"
+    "- Answer in the same language as the question."
+)
+
+_NO_PASSAGES_ANSWER = "I couldn't find anything relevant to that question in your stored documents."
+
+
+class _AskOutput(ApiModel):
+    """Raw model answer for the single ask pass."""
+
+    answer: str = Field(description="The answer grounded in the passages, with inline citations.")
+
+
+def _meta_int(value: str | None) -> int | None:
+    if value is None:
+        return None
+    try:
+        return int(value)
+    except ValueError:
+        return None
+
+
+def passage_from_hit(hit: CollectionSearchHit) -> DocumentPassage:
+    """Map a store search hit onto the wire passage shape.
+
+    Docparse chunks carry page bounds and a heading path; plain page-text
+    chunks only carry ``page_number``, which maps to both bounds.
+    """
+    meta = hit.result.document.metadata
+    page_start = _meta_int(meta.get(_PAGE_START_KEY))
+    page_end = _meta_int(meta.get(_PAGE_END_KEY))
+    if page_start is None and page_end is None:
+        page_start = page_end = _meta_int(meta.get(PAGE_NUMBER_METADATA_KEY))
+    heading = meta.get(_HEADING_PATH_KEY)
+    source = meta.get("source")
+    if source and ":page:" in source:
+        # Page-text chunk sources look like "report.pdf:page:3"; show the file name.
+        source = source.rsplit(":page:", 1)[0]
+    return DocumentPassage(
+        document_id=hit.collection,
+        text=hit.result.document.text,
+        score=hit.result.score,
+        page_start=page_start,
+        page_end=page_end,
+        heading_path=heading.split(_HEADING_PATH_SEPARATOR) if heading else [],
+        source=source or None,
+    )
+
+
+def format_passages(passages: list[DocumentPassage]) -> str:
+    """Render passages for the prompt with the citation handle in each header."""
+    return "\n\n".join(_format_passage(i, passage) for i, passage in enumerate(passages, 1))
+
+
+def _format_passage(index: int, passage: DocumentPassage) -> str:
+    name = passage.source or passage.document_id
+    if passage.page_start is None:
+        pages = ""
+    elif passage.page_end is not None and passage.page_end != passage.page_start:
+        pages = f" p.{passage.page_start}-{passage.page_end}"
+    else:
+        pages = f" p.{passage.page_start}"
+    return f"[Passage {index} | {name}{pages}]\n{passage.text}"
+
+
+class KnowledgeAskAgent:
+    """Answers a question from the caller's stored documents.
+
+    Retrieves the top passages the caller can read (same path as the search
+    endpoint), then runs one smart-model pass over just those passages.
+    """
+
+    def __init__(self, runtime: AppRuntime) -> None:
+        self.runtime = runtime
+        # Ollama/custom block tool-calling under native json-schema output; see agents.output_mode.
+        provider = runtime.settings.chat_provider
+        self._agent: Agent[None, _AskOutput] = Agent(
+            model=runtime.smart_model,
+            output_type=structured_output([_AskOutput], chat_provider=provider),
+            system_prompt=_SYSTEM_PROMPT,
+            model_settings=runtime.smart_model_settings,
+            retries=output_retries(provider),
+        )
+
+    async def ask(self, request: AskDocumentsRequest, principals: list[PrincipalId]) -> AskDocumentsResponse:
+        hits = await self.runtime.documents.search_with_collections(
+            request.question, principals=principals, top_k=request.top_k
+        )
+        passages = [passage_from_hit(hit) for hit in hits]
+        if not passages:
+            logger.info("[knowledge-ask] question=%r -> 0 passages", request.question)
+            return AskDocumentsResponse(answer=_NO_PASSAGES_ANSWER, passages=[])
+        prompt = f"Question: {request.question}\n\nPassages:\n{format_passages(passages)}"
+        logger.debug("[knowledge-ask] prompt:\n%s", prompt)
+        result = await self._agent.run(prompt)
+        return AskDocumentsResponse(answer=result.output.answer, passages=passages)

engine/src/stirling/api/bootstrap.py

@@ -10,6 +10,7 @@ from pydantic_ai.models import Model
 from stirling.agents import (
     DocumentClassifierAgent,
     ExecutionPlanningAgent,
+    KnowledgeAskAgent,
     OrchestratorAgent,
     PdfEditAgent,
     PdfQuestionAgent,
@@ -18,6 +19,7 @@ from stirling.agents import (
 from stirling.agents.ledger import MathAuditorAgent
 from stirling.agents.pdf_comment import PdfCommentAgent
 from stirling.config import AppSettings
+from stirling.docparse import ExtractFieldsAgent, SmartSplitAgent, SuggestSchemaAgent
 from stirling.documents import DocumentService, EmbeddingService
 from stirling.services import AppRuntime, build_runtime
 
@@ -35,6 +37,10 @@ class AppState:
     math_auditor_agent: MathAuditorAgent
     pdf_comment_agent: PdfCommentAgent
     document_classifier_agent: DocumentClassifierAgent
+    knowledge_ask_agent: KnowledgeAskAgent
+    extract_fields_agent: ExtractFieldsAgent
+    smart_split_agent: SmartSplitAgent
+    suggest_schema_agent: SuggestSchemaAgent
 
 
 def build_app_state(
@@ -63,6 +69,10 @@ def build_app_state(
         math_auditor_agent=MathAuditorAgent(runtime),
         pdf_comment_agent=PdfCommentAgent(runtime),
         document_classifier_agent=DocumentClassifierAgent(runtime),
+        knowledge_ask_agent=KnowledgeAskAgent(runtime),
+        extract_fields_agent=ExtractFieldsAgent(runtime),
+        smart_split_agent=SmartSplitAgent(runtime),
+        suggest_schema_agent=SuggestSchemaAgent(runtime),
     )
 
 

engine/src/stirling/api/dependencies.py

@@ -7,6 +7,7 @@ from fastapi import Depends, HTTPException, Request, status
 from stirling.agents import (
     DocumentClassifierAgent,
     ExecutionPlanningAgent,
+    KnowledgeAskAgent,
     OrchestratorAgent,
     PdfEditAgent,
     PdfQuestionAgent,
@@ -15,6 +16,7 @@ from stirling.agents import (
 from stirling.agents.ledger import MathAuditorAgent
 from stirling.agents.pdf_comment import PdfCommentAgent
 from stirling.config import AppSettings, load_settings
+from stirling.docparse import ExtractFieldsAgent, SmartSplitAgent, SuggestSchemaAgent
 from stirling.documents import DocumentService
 from stirling.models import UserId
 from stirling.services import AppRuntime, current_user_id
@@ -60,6 +62,22 @@ def get_document_classifier_agent(request: Request) -> DocumentClassifierAgent:
     return request.app.state.document_classifier_agent
 
 
+def get_knowledge_ask_agent(request: Request) -> KnowledgeAskAgent:
+    return request.app.state.knowledge_ask_agent
+
+
+def get_extract_fields_agent(request: Request) -> ExtractFieldsAgent:
+    return request.app.state.extract_fields_agent
+
+
+def get_smart_split_agent(request: Request) -> SmartSplitAgent:
+    return request.app.state.smart_split_agent
+
+
+def get_suggest_schema_agent(request: Request) -> SuggestSchemaAgent:
+    return request.app.state.suggest_schema_agent
+
+
 def require_user_id() -> UserId:
     """FastAPI dependency for routes that touch per-user storage.
 

engine/src/stirling/api/routes/docparse.py

@@ -1,9 +1,9 @@
-"""DocParse routes: parse, tables, rag-ingest, capabilities.
+"""DocParse routes: parse, extract, split, chunk, tables, fill, capabilities.
 
-Tier routing: requests carrying raw file bytes can use the advanced (Docling)
-path when the addon is installed; text-only requests run the basic path.
-Forcing ``advanced`` without the addon returns 501 with a machine-readable
-``addonRequired`` detail that Java maps onto its own error.
+Tier routing happens here: requests carrying raw file bytes can use the
+advanced (Docling) path when the addon is installed; text-only requests run
+the basic path. Forcing ``advanced`` without the addon returns 501 with a
+machine-readable ``addonRequired`` detail that Java maps onto its own error.
 """
 
 from __future__ import annotations
@@ -16,22 +16,42 @@ from typing import Annotated
 import anyio.to_thread
 from fastapi import APIRouter, Depends, HTTPException, status
 
-from stirling.api.dependencies import get_document_service, require_user_id
+from stirling.api.dependencies import (
+    get_document_service,
+    get_extract_fields_agent,
+    get_smart_split_agent,
+    get_suggest_schema_agent,
+    require_user_id,
+)
 from stirling.config import AppSettings, load_settings
 from stirling.contracts.docparse import (
+    ChunkDocumentRequest,
+    ChunkDocumentResponse,
     DocChunk,
     DocparseCapabilities,
     DocparseMode,
+    DocparseTier,
+    ExtractFieldsRequest,
+    ExtractFieldsResponse,
     ExtractTablesRequest,
     ExtractTablesResponse,
+    FillDocxRequest,
+    FillDocxResponse,
     ParseDocumentRequest,
     ParseDocumentResponse,
     RagIngestRequest,
     RagIngestResponse,
+    SmartSplitRequest,
+    SmartSplitResponse,
+    SuggestSchemaRequest,
+    SuggestSchemaResponse,
 )
-from stirling.docparse import basic_chunks, probe_capabilities
+from stirling.docparse import basic_chunks, fill_docx, probe_capabilities
 from stirling.docparse.capability import models_dir
 from stirling.docparse.chunking import advanced_chunks
+from stirling.docparse.extractor import ExtractFieldsAgent, SchemaError, pages_from_parse
+from stirling.docparse.splitter import SmartSplitAgent
+from stirling.docparse.suggest_schema import SuggestSchemaAgent
 from stirling.documents import DocumentService
 from stirling.documents.service import CONTENT_TYPE_METADATA_KEY, DOCPARSE_CHUNK_CONTENT_TYPE
 from stirling.models import OwnerId, PrincipalId, UserId
@@ -101,6 +121,96 @@ async def parse_document(request: ParseDocumentRequest) -> ParseDocumentResponse
     )
 
 
+@router.post("/extract", response_model=ExtractFieldsResponse)
+async def extract_fields(
+    request: ExtractFieldsRequest,
+    agent: Annotated[ExtractFieldsAgent, Depends(get_extract_fields_agent)],
+) -> ExtractFieldsResponse:
+    settings = _settings()
+    caps = _capabilities(settings)
+
+    use_advanced = request.mode is DocparseMode.ADVANCED or (
+        request.mode is DocparseMode.AUTO and caps.advanced_installed and request.content_base64 is not None
+    )
+    parse = None
+    if use_advanced:
+        artifacts = _require_advanced(settings)
+        if request.content_base64 is None:
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail="advanced extraction needs contentBase64 (the raw file)",
+            )
+        parse = await _parse_advanced(request.content_base64, request.file_name, with_ocr=True, artifacts=artifacts)
+
+    pages = request.pages or (pages_from_parse(parse) if parse is not None else None)
+    if not pages:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+            detail="send pages (extracted text) or contentBase64 with the addon installed",
+        )
+    try:
+        return await agent.extract(request, pages, parse)
+    except SchemaError as error:
+        raise HTTPException(status_code=status.HTTP_422_UNPROCESSABLE_ENTITY, detail=str(error)) from error
+
+
+@router.post("/suggest-schema", response_model=SuggestSchemaResponse)
+async def suggest_schema(
+    request: SuggestSchemaRequest,
+    agent: Annotated[SuggestSchemaAgent, Depends(get_suggest_schema_agent)],
+) -> SuggestSchemaResponse:
+    """Propose an extraction schema from the document's first pages.
+    Tier routing: pages -> basic; contentBase64 + addon -> advanced parse."""
+    settings = _settings()
+    caps = _capabilities(settings)
+    pages = request.pages
+    tier = DocparseTier.BASIC
+    if not pages and request.content_base64 is not None and caps.advanced_installed:
+        artifacts = _require_advanced(settings)
+        parse = await _parse_advanced(request.content_base64, request.file_name, with_ocr=True, artifacts=artifacts)
+        pages = pages_from_parse(parse)
+        tier = DocparseTier.ADVANCED
+    if not pages:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+            detail="send pages (extracted text) or contentBase64 with the addon installed",
+        )
+    return await agent.suggest(request, pages, tier)
+
+
+@router.post("/split", response_model=SmartSplitResponse)
+async def smart_split(
+    request: SmartSplitRequest,
+    agent: Annotated[SmartSplitAgent, Depends(get_smart_split_agent)],
+) -> SmartSplitResponse:
+    return await agent.split(request)
+
+
+@router.post("/chunk", response_model=ChunkDocumentResponse)
+async def chunk_document(request: ChunkDocumentRequest) -> ChunkDocumentResponse:
+    settings = _settings()
+    caps = _capabilities(settings)
+    use_advanced = request.mode is DocparseMode.ADVANCED or (
+        request.mode is DocparseMode.AUTO and caps.advanced_installed and request.content_base64 is not None
+    )
+    if use_advanced:
+        artifacts = _require_advanced(settings)
+        if request.content_base64 is None:
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail="advanced chunking needs contentBase64 (the raw file)",
+            )
+        parse = await _parse_advanced(request.content_base64, request.file_name, with_ocr=True, artifacts=artifacts)
+        return advanced_chunks(parse, request.chunk_size, request.overlap)
+
+    if not request.pages:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+            detail="send pages (extracted text) or contentBase64 with the addon installed",
+        )
+    return basic_chunks(request.pages, request.chunk_size, request.overlap)
+
+
 def _chunk_metadata(chunk: DocChunk) -> dict[str, str]:
     meta = {CONTENT_TYPE_METADATA_KEY: DOCPARSE_CHUNK_CONTENT_TYPE}
     if chunk.page_start is not None:
@@ -192,3 +302,13 @@ async def extract_tables(request: ExtractTablesRequest) -> ExtractTablesResponse
     artifacts = _require_advanced(settings)
     parse = await _parse_advanced(request.content_base64, request.file_name, with_ocr=True, artifacts=artifacts)
     return ExtractTablesResponse(mode=parse.mode, tables=parse.tables)
+
+
+@router.post("/fill-docx", response_model=FillDocxResponse)
+async def fill_docx_template(request: FillDocxRequest) -> FillDocxResponse:
+    try:
+        return await anyio.to_thread.run_sync(lambda: fill_docx(request))
+    except (KeyError, ValueError) as error:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY, detail=f"invalid docx template: {error}"
+        ) from error

engine/src/stirling/api/routes/documents.py

@@ -5,13 +5,24 @@ from typing import Annotated
 
 from fastapi import APIRouter, Depends
 
-from stirling.api.dependencies import get_document_service, require_user_id
+from stirling.agents.knowledge_ask import KnowledgeAskAgent, passage_from_hit
+from stirling.api.dependencies import get_document_service, get_knowledge_ask_agent, require_user_id
+from stirling.config import load_settings
 from stirling.contracts import (
     DeleteDocumentResponse,
     IngestDocumentRequest,
     IngestDocumentResponse,
 )
-from stirling.contracts.documents import PurgeOwnerResponse
+from stirling.contracts.documents import (
+    AskDocumentsRequest,
+    AskDocumentsResponse,
+    DocumentStatsResponse,
+    DocumentSummary,
+    ListDocumentsResponse,
+    PurgeOwnerResponse,
+    SearchDocumentsRequest,
+    SearchDocumentsResponse,
+)
 from stirling.documents import DocumentService
 from stirling.models import FileId, OwnerId, PrincipalId, UserId
 
@@ -45,6 +56,70 @@ async def ingest_document(
     return IngestDocumentResponse(document_id=request.document_id, chunks_indexed=chunks_indexed)
 
 
+@router.get("/stats", response_model=DocumentStatsResponse)
+async def document_stats(
+    documents: Annotated[DocumentService, Depends(get_document_service)],
+    _user_id: Annotated[UserId, Depends(require_user_id)],
+) -> DocumentStatsResponse:
+    """Deployment-wide store counts for the admin dashboard.
+
+    Not tenant-filtered: counts cover every owner's content, so this reports
+    what the whole store holds, not what the caller can read.
+    """
+    settings = load_settings()
+    counts = await documents.stats()
+    return DocumentStatsResponse(
+        backend=settings.documents_backend.value,
+        documents=counts.documents,
+        chunks=counts.chunks,
+        embedding_model=settings.rag_embedding_model,
+    )
+
+
+@router.get("/list", response_model=ListDocumentsResponse)
+async def list_documents(
+    documents: Annotated[DocumentService, Depends(get_document_service)],
+    user_id: Annotated[UserId, Depends(require_user_id)],
+) -> ListDocumentsResponse:
+    """Per-document rollup of what the caller can read: distinct document ids
+    with their stored source label and chunk count. Never shows another
+    principal's documents.
+    """
+    summaries = await documents.list_documents([PrincipalId(user_id)])
+    return ListDocumentsResponse(
+        documents=[
+            DocumentSummary(document_id=FileId(s.collection), source=s.source, chunks=s.chunks) for s in summaries
+        ]
+    )
+
+
+@router.post("/search", response_model=SearchDocumentsResponse)
+async def search_documents(
+    request: SearchDocumentsRequest,
+    documents: Annotated[DocumentService, Depends(get_document_service)],
+    user_id: Annotated[UserId, Depends(require_user_id)],
+) -> SearchDocumentsResponse:
+    """Semantic search across every document the caller can read.
+
+    Same retrieval path as the RAG toolset: embed the query, search the
+    caller's readable collections, merge by score.
+    """
+    hits = await documents.search_with_collections(
+        request.query, principals=[PrincipalId(user_id)], top_k=request.top_k
+    )
+    return SearchDocumentsResponse(passages=[passage_from_hit(hit) for hit in hits])
+
+
+@router.post("/ask", response_model=AskDocumentsResponse)
+async def ask_documents(
+    request: AskDocumentsRequest,
+    agent: Annotated[KnowledgeAskAgent, Depends(get_knowledge_ask_agent)],
+    user_id: Annotated[UserId, Depends(require_user_id)],
+) -> AskDocumentsResponse:
+    """Answer a question from the caller's stored documents with inline citations."""
+    return await agent.ask(request, principals=[PrincipalId(user_id)])
+
+
 @router.delete("/by-id/{document_id}", response_model=DeleteDocumentResponse)
 async def delete_document(
     document_id: FileId,

engine/src/stirling/contracts/__init__.py

@@ -42,6 +42,36 @@ from .contradiction import (
     ContradictionReport,
     ContradictionSeverity,
 )
+from .docparse import (
+    BlockType,
+    ChunkDocumentRequest,
+    ChunkDocumentResponse,
+    DocBlock,
+    DocChunk,
+    DocparseCapabilities,
+    DocparseMode,
+    DocparseTier,
+    DocTable,
+    ExtractedField,
+    ExtractFieldsRequest,
+    ExtractFieldsResponse,
+    ExtractTablesRequest,
+    ExtractTablesResponse,
+    FieldCitation,
+    FillDocxRequest,
+    FillDocxResponse,
+    ParseDocumentRequest,
+    ParseDocumentResponse,
+    RagIngestRequest,
+    RagIngestResponse,
+    SmartSplitRequest,
+    SmartSplitResponse,
+    SplitPart,
+    SuggestedField,
+    SuggestedFieldType,
+    SuggestSchemaRequest,
+    SuggestSchemaResponse,
+)
 from .document_classifier import (
     ClassifyDocumentRequest,
     ClassifyDocumentResponse,
@@ -49,13 +79,21 @@ from .document_classifier import (
     LabelOption,
 )
 from .documents import (
+    AskDocumentsRequest,
+    AskDocumentsResponse,
     DeleteDocumentResponse,
+    DocumentPassage,
+    DocumentStatsResponse,
+    DocumentSummary,
     IngestDocumentRequest,
     IngestDocumentResponse,
+    ListDocumentsResponse,
     Page,
     PageRange,
     PageText,
     PurgeOwnerResponse,
+    SearchDocumentsRequest,
+    SearchDocumentsResponse,
 )
 from .execution import (
     AgentExecutionRequest,
@@ -140,10 +178,15 @@ __all__ = [
     "AiFile",
     "AiToolAgentStep",
     "ArtifactKind",
+    "AskDocumentsRequest",
+    "AskDocumentsResponse",
+    "BlockType",
     "CannotContinueExecutionAction",
+    "ChunkDocumentRequest",
+    "ChunkDocumentResponse",
+    "Claim",
     "ClassifyDocumentRequest",
     "ClassifyDocumentResponse",
-    "Claim",
     "CommentSpec",
     "CompletedExecutionAction",
     "ConfigApplyResponse",
@@ -156,30 +199,45 @@ __all__ = [
     "ContradictionSeverity",
     "ConversationMessage",
     "DeleteDocumentResponse",
-    "PurgeOwnerResponse",
     "Discrepancy",
-    "DocumentClassificationResponse",
-    "LabelOption",
-    "DocumentMeta",
-    "DocumentSections",
     "DiscrepancyKind",
+    "DocBlock",
+    "DocChunk",
+    "DocTable",
+    "DocparseCapabilities",
+    "DocparseMode",
+    "DocparseTier",
+    "DocumentClassificationResponse",
+    "DocumentMeta",
+    "DocumentPassage",
+    "DocumentSections",
+    "DocumentStatsResponse",
+    "DocumentSummary",
     "EditCannotDoResponse",
     "EditClarificationRequest",
     "EditPlanResponse",
     "Evidence",
     "ExecutionContext",
     "ExecutionStepResult",
+    "ExtractFieldsRequest",
+    "ExtractFieldsResponse",
+    "ExtractTablesRequest",
+    "ExtractTablesResponse",
+    "ExtractedField",
     "ExtractedFileText",
     "ExtractedTextArtifact",
+    "FieldCitation",
+    "FillDocxRequest",
+    "FillDocxResponse",
     "Folio",
     "FolioManifest",
     "FolioType",
-    "format_conversation_history",
-    "format_file_names",
     "GenerateFileResponse",
     "HealthResponse",
     "IngestDocumentRequest",
     "IngestDocumentResponse",
+    "LabelOption",
+    "ListDocumentsResponse",
     "MathAuditorToolReportArtifact",
     "NeedContentFileRequest",
     "NeedContentResponse",
@@ -190,6 +248,8 @@ __all__ = [
     "Page",
     "PageRange",
     "PageText",
+    "ParseDocumentRequest",
+    "ParseDocumentResponse",
     "PdfCommentInstruction",
     "PdfCommentReport",
     "PdfCommentRequest",
@@ -212,9 +272,21 @@ __all__ = [
     "PdfReviewOrchestrateResponse",
     "PdfTextSelection",
     "ProgressEvent",
+    "PurgeOwnerResponse",
+    "RagIngestRequest",
+    "RagIngestResponse",
     "Requisition",
+    "SearchDocumentsRequest",
+    "SearchDocumentsResponse",
     "Severity",
+    "SmartSplitRequest",
+    "SmartSplitResponse",
+    "SplitPart",
     "StepKind",
+    "SuggestSchemaRequest",
+    "SuggestSchemaResponse",
+    "SuggestedField",
+    "SuggestedFieldType",
     "SupportedCapability",
     "TextChunk",
     "ToolCallExecutionAction",
@@ -228,4 +300,6 @@ __all__ = [
     "WholeDocSliceDone",
     "WorkflowArtifact",
     "WorkflowOutcome",
+    "format_conversation_history",
+    "format_file_names",
 ]

engine/src/stirling/contracts/docparse.py

@@ -10,7 +10,7 @@ from __future__ import annotations
 from datetime import datetime
 from enum import StrEnum
 
-from pydantic import Field
+from pydantic import Field, JsonValue
 
 from stirling.contracts.documents import PageText
 from stirling.models import ApiModel, FileId, OwnerId, PrincipalId
@@ -140,6 +140,112 @@ class RagIngestResponse(ApiModel):
     chunks: list[DocChunk] | None = None
 
 
+class FieldCitation(ApiModel):
+    """Where a value came from. ``quote`` is always set; ``bbox`` only when a
+    layout parse ran (advanced tier); offsets index into the cited page's text."""
+
+    page: int | None = Field(default=None, ge=1)
+    bbox: list[float] | None = None
+    quote: str
+    start_offset: int | None = Field(default=None, ge=0)
+    end_offset: int | None = Field(default=None, ge=0)
+
+
+class ExtractedField(ApiModel):
+    name: str
+    value: JsonValue = None
+    confidence: float = Field(ge=0.0, le=1.0)
+    citations: list[FieldCitation] = Field(default_factory=list)
+
+
+class ExtractFieldsRequest(ApiModel):
+    """``pages`` drives the basic tier (caller-extracted text); ``content_base64``
+    lets the advanced tier parse the raw file itself. Send either or both."""
+
+    file_name: str = Field(min_length=1)
+    fields_schema: dict[str, JsonValue]
+    pages: list[PageText] | None = None
+    content_base64: str | None = None
+    mode: DocparseMode = DocparseMode.AUTO
+    instructions: str | None = None
+
+
+class ExtractFieldsResponse(ApiModel):
+    mode: DocparseTier
+    fields: list[ExtractedField] = Field(default_factory=list)
+    overall_confidence: float = Field(ge=0.0, le=1.0)
+
+
+class SuggestedFieldType(StrEnum):
+    """Scalar types the schema suggester may propose; the extractor's leaf subset."""
+
+    STRING = "string"
+    NUMBER = "number"
+    INTEGER = "integer"
+    BOOLEAN = "boolean"
+
+
+class SuggestedField(ApiModel):
+    name: str = Field(description="snake_case field identifier, e.g. 'invoice_number'.")
+    type: SuggestedFieldType
+    description: str = ""
+
+
+class SuggestSchemaRequest(ApiModel):
+    """``pages`` drives the basic tier (caller-extracted text); ``content_base64``
+    lets the advanced tier parse the raw file itself. Send either."""
+
+    file_name: str = Field(min_length=1)
+    pages: list[PageText] | None = None
+    content_base64: str | None = None
+    max_fields: int = Field(default=8, ge=1, le=20)
+
+
+class SuggestSchemaResponse(ApiModel):
+    mode: DocparseTier
+    fields: list[SuggestedField] = Field(default_factory=list)
+
+
+class SplitPart(ApiModel):
+    start_page: int = Field(ge=1)
+    end_page: int = Field(ge=1)
+    label: str
+    confidence: float = Field(ge=0.0, le=1.0)
+
+
+class SmartSplitRequest(ApiModel):
+    file_name: str = Field(min_length=1)
+    rule: str = Field(
+        min_length=1, description="Natural-language boundary rule, e.g. 'split where a new invoice starts'."
+    )
+    pages: list[PageText]
+    max_parts: int = Field(default=50, ge=1, le=500)
+
+
+class SmartSplitResponse(ApiModel):
+    parts: list[SplitPart] = Field(default_factory=list)
+
+
+class ChunkDocumentRequest(ApiModel):
+    file_name: str = Field(min_length=1)
+    pages: list[PageText] | None = None
+    content_base64: str | None = None
+    chunk_size: int = Field(default=512, ge=64, le=32_768)
+    overlap: int = Field(default=64, ge=0, le=4_096)
+    mode: DocparseMode = DocparseMode.AUTO
+
+
+class FillDocxRequest(ApiModel):
+    template_base64: str = Field(min_length=1)
+    data: dict[str, JsonValue]
+
+
+class FillDocxResponse(ApiModel):
+    docx_base64: str
+    replaced: int = Field(ge=0)
+    missing: list[str] = Field(default_factory=list)
+
+
 class DocparseCapabilities(ApiModel):
     """What the engine can actually do right now; Java caches and republishes this."""
 

engine/src/stirling/contracts/documents.py

@@ -75,3 +75,65 @@ class PurgeOwnerResponse(ApiModel):
 
     owner_id: OwnerId
     deleted: int = Field(ge=0)
+
+
+class DocumentStatsResponse(ApiModel):
+    """Returned by ``GET /api/v1/documents/stats``. Deployment-wide counts
+    (every owner's content) powering the admin dashboard."""
+
+    backend: str
+    documents: int = Field(ge=0)
+    chunks: int = Field(ge=0)
+    embedding_model: str
+
+
+class DocumentSummary(ApiModel):
+    """One stored document the caller can read: its id, source label, chunk count."""
+
+    document_id: FileId
+    source: str
+    chunks: int = Field(ge=0)
+
+
+class ListDocumentsResponse(ApiModel):
+    """Returned by ``GET /api/v1/documents/list``. Caller-scoped rollup."""
+
+    documents: list[DocumentSummary]
+
+
+class SearchDocumentsRequest(ApiModel):
+    """Semantic search over every document the caller can read."""
+
+    query: str = Field(min_length=1)
+    top_k: int = Field(default=8, ge=1, le=50)
+
+
+class DocumentPassage(ApiModel):
+    """A retrieved chunk on the wire. Page bounds and heading path come from
+    chunk metadata when present (docparse chunks carry them); nulls otherwise."""
+
+    document_id: FileId
+    text: str
+    score: float
+    page_start: int | None = None
+    page_end: int | None = None
+    heading_path: list[str] = Field(default_factory=list)
+    source: str | None = None
+
+
+class SearchDocumentsResponse(ApiModel):
+    passages: list[DocumentPassage]
+
+
+class AskDocumentsRequest(ApiModel):
+    """Question answered only from the caller's stored documents."""
+
+    question: str = Field(min_length=1)
+    top_k: int = Field(default=8, ge=1, le=20)
+
+
+class AskDocumentsResponse(ApiModel):
+    """Grounded answer with inline citations plus the passages it drew from."""
+
+    answer: str
+    passages: list[DocumentPassage]

engine/src/stirling/docparse/__init__.py

@@ -8,10 +8,18 @@ from __future__ import annotations
 
 from stirling.docparse.capability import activate_site, probe_capabilities
 from stirling.docparse.chunking import advanced_chunks, basic_chunks
+from stirling.docparse.docxfill import fill_docx
+from stirling.docparse.extractor import ExtractFieldsAgent
+from stirling.docparse.splitter import SmartSplitAgent
+from stirling.docparse.suggest_schema import SuggestSchemaAgent
 
 __all__ = [
+    "ExtractFieldsAgent",
+    "SmartSplitAgent",
+    "SuggestSchemaAgent",
     "activate_site",
     "advanced_chunks",
     "basic_chunks",
+    "fill_docx",
     "probe_capabilities",
 ]

engine/src/stirling/docparse/docxfill.py

@@ -0,0 +1,178 @@
+"""Fill DOCX templates from JSON data: ``{{ dotted.path }}`` placeholders.
+
+Scalar placeholders are replaced everywhere (body, tables, headers, footers).
+A table row whose text contains ``{{#items.field}}`` markers is treated as a
+row template: it is cloned once per element of the ``items`` array. Unresolved
+placeholders are left in place and reported back so the caller can surface them.
+
+Formatting caveat: a placeholder split across styled runs collapses that
+paragraph's text into its first run's style.
+"""
+
+from __future__ import annotations
+
+import base64
+import copy
+import io
+import re
+from typing import Any
+
+from pydantic import JsonValue
+
+from stirling.contracts.docparse import FillDocxRequest, FillDocxResponse
+
+_PLACEHOLDER = re.compile(r"\{\{\s*(#?[\w.]+)\s*\}\}")
+
+
+def _resolve(path: str, data: dict[str, Any]) -> Any | None:
+    node: Any = data
+    for part in path.split("."):
+        if isinstance(node, dict) and part in node:
+            node = node[part]
+        else:
+            return None
+    return node
+
+
+def _render_scalar(value: Any) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, bool):
+        return "true" if value else "false"
+    if isinstance(value, list):
+        return ", ".join(_render_scalar(v) for v in value)
+    return str(value)
+
+
+class _Stats:
+    def __init__(self) -> None:
+        self.replaced = 0
+        self.missing: set[str] = set()
+
+
+def _fill_paragraph(paragraph: Any, data: dict[str, Any], stats: _Stats) -> None:
+    text = paragraph.text
+    if "{{" not in text:
+        return
+
+    def substitute(match: re.Match[str]) -> str:
+        path = match.group(1)
+        if path.startswith("#"):
+            return match.group(0)  # row-template marker, handled at table level
+        value = _resolve(path, data)
+        if value is None:
+            stats.missing.add(path)
+            return match.group(0)
+        stats.replaced += 1
+        return _render_scalar(value)
+
+    rendered = _PLACEHOLDER.sub(substitute, text)
+    if rendered == text:
+        return
+    # Collapse into the first run to survive placeholders split across runs.
+    if paragraph.runs:
+        paragraph.runs[0].text = rendered
+        for run in paragraph.runs[1:]:
+            run.text = ""
+    else:
+        paragraph.add_run(rendered)
+
+
+def _row_template_array(row: Any) -> str | None:
+    """Return the array name when the row carries ``{{#name.field}}`` markers."""
+    names = {
+        match.group(1)[1:].split(".")[0]
+        for cell in row.cells
+        for match in _PLACEHOLDER.finditer(cell.text)
+        if match.group(1).startswith("#")
+    }
+    return names.pop() if len(names) == 1 else None
+
+
+def _fill_table(table: Any, data: dict[str, Any], stats: _Stats) -> None:
+    for row in list(table.rows):
+        array_name = _row_template_array(row)
+        if array_name is None:
+            continue
+        items = _resolve(array_name, data)
+        if not isinstance(items, list):
+            stats.missing.add(array_name)
+            continue
+        for _ in items:
+            new_row = copy.deepcopy(row._tr)
+            row._tr.addprevious(new_row)
+        # Clones sit before the template; rewrite their markers, then drop the template.
+        _rewrite_cloned_rows(table, row, array_name, items, data, stats)
+        row._tr.getparent().remove(row._tr)
+
+
+def _rewrite_cloned_rows(
+    table: Any, template_row: Any, array_name: str, items: list[Any], data: dict[str, Any], stats: _Stats
+) -> None:
+    marker_prefix = f"#{array_name}"
+    clones = [
+        r for r in table.rows if r._tr is not template_row._tr and marker_prefix in "".join(c.text for c in r.cells)
+    ]
+    for row, item in zip(clones, items, strict=False):
+        scoped = dict(data)
+        scoped[array_name] = item if isinstance(item, dict) else {"value": item}
+        for cell in row.cells:
+            for paragraph in cell.paragraphs:
+                text = paragraph.text
+
+                def substitute(match: re.Match[str]) -> str:
+                    path = match.group(1)
+                    if not path.startswith(marker_prefix):
+                        return match.group(0)
+                    item_path = path[1:]  # "#items.field" -> "items.field"
+                    value = _resolve(item_path, scoped)
+                    if value is None and "." not in item_path:
+                        value = scoped[array_name].get("value") if isinstance(scoped[array_name], dict) else None
+                    if value is None:
+                        stats.missing.add(item_path)
+                        return match.group(0)
+                    stats.replaced += 1
+                    return _render_scalar(value)
+
+                rendered = _PLACEHOLDER.sub(substitute, text)
+                if rendered != text:
+                    if paragraph.runs:
+                        paragraph.runs[0].text = rendered
+                        for run in paragraph.runs[1:]:
+                            run.text = ""
+                    else:
+                        paragraph.add_run(rendered)
+
+
+def _walk_paragraphs(document: Any) -> list[tuple[Any, Any]]:
+    """Yield (paragraph, containing table or None) across body, tables, headers, footers."""
+    found: list[tuple[Any, Any]] = [(p, None) for p in document.paragraphs]
+    for table in document.tables:
+        for row in table.rows:
+            for cell in row.cells:
+                found.extend((p, table) for p in cell.paragraphs)
+    for section in document.sections:
+        for part in (section.header, section.footer):
+            found.extend((p, None) for p in part.paragraphs)
+    return found
+
+
+def fill_docx(request: FillDocxRequest) -> FillDocxResponse:
+    import docx  # local import: python-docx is small but only needed here
+
+    data: dict[str, JsonValue] = dict(request.data)
+    document = docx.Document(io.BytesIO(base64.b64decode(request.template_base64)))
+    stats = _Stats()
+
+    for table in document.tables:
+        _fill_table(table, data, stats)
+    for paragraph, _table in _walk_paragraphs(document):
+        _fill_paragraph(paragraph, data, stats)
+
+    out = io.BytesIO()
+    document.save(out)
+    return FillDocxResponse(
+        docx_base64=base64.b64encode(out.getvalue()).decode("ascii"),
+        replaced=stats.replaced,
+        missing=sorted(stats.missing),
+    )

engine/src/stirling/docparse/evals.py

@@ -0,0 +1,146 @@
+"""Extraction accuracy harness: score a gold-labelled case set against a live engine.
+
+Case layout (one directory per document):
+
+    cases/
+      invoice-001/
+        input.pdf            # the document
+        expected.json        # {"fieldsSchema": {...}, "fields": {"invoice_number": "INV-1", ...},
+                             #  "pages": [{"pageNumber": 1, "text": "..."}]?  (optional, for basic tier)}
+
+Run:
+    uv run python -m stirling.docparse.evals cases/ --engine http://localhost:5001 --output report.json
+
+Scoring per field: exact match, then normalized match (case/whitespace/currency
+punctuation collapsed, numeric tolerance 1e-6). The report aggregates per-case
+and overall accuracy so pipeline changes can be regression-tracked. Stdlib
+HTTP only - the harness must run anywhere the engine runs.
+"""
+
+from __future__ import annotations
+
+import argparse
+import base64
+import json
+import re
+import sys
+import urllib.error
+import urllib.request
+from pathlib import Path
+from typing import Any
+
+_NORMALIZE_STRIP = re.compile(r"[\s,€$£%]+")
+
+
+def normalized_equal(expected: Any, actual: Any) -> bool:
+    if expected is None or actual is None:
+        return expected is actual
+    try:
+        return abs(float(expected) - float(actual)) < 1e-6
+    except (TypeError, ValueError):
+        pass
+    if isinstance(expected, list) and isinstance(actual, list):
+        return len(expected) == len(actual) and all(normalized_equal(e, a) for e, a in zip(expected, actual))
+    return _NORMALIZE_STRIP.sub("", str(expected)).casefold() == _NORMALIZE_STRIP.sub("", str(actual)).casefold()
+
+
+def _call_extract(engine: str, secret: str | None, payload: dict[str, Any]) -> dict[str, Any]:
+    request = urllib.request.Request(
+        f"{engine.rstrip('/')}/api/v1/docparse/extract",
+        data=json.dumps(payload).encode("utf-8"),
+        headers={"Content-Type": "application/json", **({"X-Engine-Auth": secret} if secret else {})},
+        method="POST",
+    )
+    with urllib.request.urlopen(request, timeout=600) as response:
+        return json.loads(response.read().decode("utf-8"))
+
+
+def score_case(engine: str, secret: str | None, case_dir: Path) -> dict[str, Any]:
+    expected_spec = json.loads((case_dir / "expected.json").read_text(encoding="utf-8"))
+    payload: dict[str, Any] = {
+        "fileName": "input.pdf",
+        "fieldsSchema": expected_spec["fieldsSchema"],
+    }
+    input_pdf = case_dir / "input.pdf"
+    if input_pdf.exists():
+        payload["contentBase64"] = base64.b64encode(input_pdf.read_bytes()).decode("ascii")
+    if expected_spec.get("pages"):
+        payload["pages"] = expected_spec["pages"]
+
+    response = _call_extract(engine, secret, payload)
+    actual = {field["name"]: field for field in response.get("fields", [])}
+
+    fields: list[dict[str, Any]] = []
+    exact = 0
+    normalized = 0
+    for name, expected_value in expected_spec.get("fields", {}).items():
+        actual_field = actual.get(name, {})
+        actual_value = actual_field.get("value")
+        is_exact = expected_value == actual_value
+        is_normalized = is_exact or normalized_equal(expected_value, actual_value)
+        exact += is_exact
+        normalized += is_normalized
+        fields.append(
+            {
+                "name": name,
+                "expected": expected_value,
+                "actual": actual_value,
+                "exact": is_exact,
+                "normalized": is_normalized,
+                "confidence": actual_field.get("confidence"),
+                "cited": bool(actual_field.get("citations")),
+            }
+        )
+
+    total = len(fields) or 1
+    return {
+        "case": case_dir.name,
+        "mode": response.get("mode"),
+        "fields": fields,
+        "exactAccuracy": round(exact / total, 4),
+        "normalizedAccuracy": round(normalized / total, 4),
+    }
+
+
+def run(cases_root: Path, engine: str, secret: str | None) -> dict[str, Any]:
+    case_dirs = sorted(d for d in cases_root.iterdir() if d.is_dir() and (d / "expected.json").exists())
+    results: list[dict[str, Any]] = []
+    failures: list[dict[str, str]] = []
+    for case_dir in case_dirs:
+        try:
+            results.append(score_case(engine, secret, case_dir))
+        except (urllib.error.URLError, OSError, KeyError, ValueError) as error:
+            failures.append({"case": case_dir.name, "error": str(error)})
+
+    scored = [r for r in results if r["fields"]]
+    overall = {
+        "cases": len(case_dirs),
+        "scored": len(scored),
+        "errors": failures,
+        "exactAccuracy": round(sum(r["exactAccuracy"] for r in scored) / len(scored), 4) if scored else 0.0,
+        "normalizedAccuracy": round(sum(r["normalizedAccuracy"] for r in scored) / len(scored), 4) if scored else 0.0,
+        "results": results,
+    }
+    return overall
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Score docparse extraction against a gold case set.")
+    parser.add_argument("cases", help="Directory of case subdirectories")
+    parser.add_argument("--engine", default="http://localhost:5001", help="Engine base URL")
+    parser.add_argument("--secret", default=None, help="X-Engine-Auth shared secret, if the engine requires it")
+    parser.add_argument("--output", default=None, help="Write the JSON report here (default: stdout)")
+    args = parser.parse_args()
+
+    report = run(Path(args.cases), args.engine, args.secret)
+    rendered = json.dumps(report, indent=2)
+    if args.output:
+        Path(args.output).write_text(rendered, encoding="utf-8")
+        print(f"exact={report['exactAccuracy']} normalized={report['normalizedAccuracy']} -> {args.output}")
+    else:
+        print(rendered)
+    return 0 if not report["errors"] else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

engine/src/stirling/docparse/extractor.py

@@ -0,0 +1,320 @@
+"""Schema-driven field extraction with grounded citations and confidence.
+
+The caller supplies a JSON Schema (subset: scalar types, enums, arrays of
+scalars, nested objects). We build a dynamic pydantic output model where every
+leaf answers ``{value, quote, confidence}``, run one smart-model pass, then
+ground each quote against the page text in code. Model confidence is damped
+when a quote can't be found - the model asserts, the grounding decides how
+much to believe it.
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from typing import Any
+
+from pydantic import BaseModel, Field, JsonValue, create_model
+from pydantic_ai import Agent
+
+from stirling.agents.output_mode import output_retries, structured_output
+from stirling.contracts.docparse import (
+    DocparseTier,
+    ExtractedField,
+    ExtractFieldsRequest,
+    ExtractFieldsResponse,
+    FieldCitation,
+    ParseDocumentResponse,
+)
+from stirling.contracts.documents import PageText
+from stirling.models import ApiModel
+from stirling.services import AppRuntime
+
+logger = logging.getLogger(__name__)
+
+# Confidence multiplier when the supporting quote can't be found in the document.
+UNGROUNDED_PENALTY = 0.6
+# Floor when the model omitted quote/confidence but the value itself is found
+# verbatim in the document - the grounding is real even if the model was terse.
+VALUE_GROUNDED_FLOOR = 0.5
+MAX_SCHEMA_DEPTH = 3
+MAX_FIELDS = 100
+
+_SYSTEM_PROMPT = (
+    "You extract structured fields from a document.\n"
+    "\n"
+    "Rules:\n"
+    "- For every field, return the value exactly as the schema types it, a short VERBATIM quote "
+    "from the document that supports it, and your confidence from 0.0 to 1.0.\n"
+    "- The quote must be copied character-for-character from the document text, at most 200 characters.\n"
+    "- If the document does not contain the field, return value null, quote null, confidence 0.0. "
+    "Never guess or fabricate.\n"
+    "- Dates: return them formatted as the schema/description asks; quote the original text.\n"
+    "- The document may be in any language."
+)
+
+
+class SchemaError(ValueError):
+    """The supplied JSON Schema is outside the supported subset."""
+
+
+def _scalar_type(spec: dict[str, Any]) -> Any:
+    # Enums stay str-typed; the allowed values travel in the field description
+    # (dynamic Literal types don't typecheck and local models handle them badly).
+    match spec.get("type"):
+        case "string":
+            return str
+        case "integer":
+            return int
+        case "number":
+            return float
+        case "boolean":
+            return bool
+        case _:
+            raise SchemaError(f"Unsupported schema type: {spec.get('type')!r}")
+
+
+def _describe(spec: dict[str, Any]) -> str | None:
+    description = spec.get("description") if isinstance(spec.get("description"), str) else None
+    enum = spec.get("enum")
+    if isinstance(enum, list) and enum:
+        allowed = ", ".join(str(v) for v in enum)
+        description = f"{description + ' ' if description else ''}Allowed values: {allowed}."
+    return description
+
+
+def _leaf_answer_model(name: str, value_type: Any, description: str | None) -> type[BaseModel]:
+    return create_model(
+        f"Answer_{re.sub(r'[^A-Za-z0-9]', '_', name)}",
+        __base__=ApiModel,
+        value=(value_type | None, Field(default=None, description=description or None)),
+        quote=(str | None, Field(default=None, max_length=400)),
+        confidence=(float, Field(default=0.0, ge=0.0, le=1.0)),
+    )
+
+
+def build_output_model(
+    fields_schema: dict[str, Any], *, _depth: int = 0, _name: str = "ExtractionOutput"
+) -> type[BaseModel]:
+    """Turn the caller's JSON Schema into a pydantic model of leaf answers."""
+    if _depth > MAX_SCHEMA_DEPTH:
+        raise SchemaError(f"Schema nesting deeper than {MAX_SCHEMA_DEPTH} is not supported")
+    properties = fields_schema.get("properties")
+    if not isinstance(properties, dict) or not properties:
+        raise SchemaError("Schema must be an object with a non-empty 'properties' map")
+    if len(properties) > MAX_FIELDS:
+        raise SchemaError(f"Schema has more than {MAX_FIELDS} fields")
+
+    model_fields: dict[str, Any] = {}
+    for raw_name, spec in properties.items():
+        name = str(raw_name)
+        if not re.fullmatch(r"[A-Za-z_][A-Za-z0-9_]*", name):
+            raise SchemaError(f"Field name {name!r} must be a valid identifier")
+        if not isinstance(spec, dict):
+            raise SchemaError(f"Field {name!r} must map to a schema object")
+        description = _describe(spec)
+
+        if spec.get("type") == "object":
+            nested = build_output_model(spec, _depth=_depth + 1, _name=f"{_name}_{name}")
+            model_fields[name] = (nested, Field(...))
+        elif spec.get("type") == "array":
+            items = spec.get("items")
+            if not isinstance(items, dict):
+                raise SchemaError(f"Array field {name!r} needs an 'items' schema")
+            if items.get("type") == "object":
+                raise SchemaError(f"Array field {name!r}: arrays of objects are not supported yet")
+            item_type: Any = _scalar_type(items)
+            answer = _leaf_answer_model(name, list[item_type], description)
+            model_fields[name] = (answer, Field(...))
+        else:
+            answer = _leaf_answer_model(name, _scalar_type(spec), description)
+            model_fields[name] = (answer, Field(...))
+
+    return create_model(_name, __base__=ApiModel, **model_fields)
+
+
+_WHITESPACE = re.compile(r"\s+")
+
+
+def _normalize(text: str) -> str:
+    return _WHITESPACE.sub(" ", text).strip().casefold()
+
+
+def find_quote(quote: str, pages: list[PageText]) -> tuple[int, int, int] | None:
+    """Locate ``quote`` in the page texts, whitespace-insensitively.
+
+    Returns (page_number, start_offset, end_offset) into the page's raw text,
+    or None. Offsets are approximate under whitespace collapsing: we search the
+    normalized page, then map back by counting non-space characters.
+    """
+    needle = _normalize(quote)
+    if not needle:
+        return None
+    for page in pages:
+        haystack = _normalize(page.text)
+        idx = haystack.find(needle)
+        if idx < 0:
+            continue
+        start = _denormalize_offset(page.text, idx)
+        end = _denormalize_offset(page.text, idx + len(needle))
+        return page.page_number, start, min(end, len(page.text))
+    return None
+
+
+def _denormalize_offset(raw: str, normalized_offset: int) -> int:
+    """Map an offset in the normalized string back into the raw string."""
+    count = 0
+    in_space = True  # leading whitespace is stripped by _normalize
+    for i, ch in enumerate(raw):
+        if ch.isspace():
+            if in_space:
+                continue
+            in_space = True
+        else:
+            in_space = False
+        if count >= normalized_offset:
+            return i
+        count += 1
+    return len(raw)
+
+
+def _bbox_for_quote(quote: str, parse: ParseDocumentResponse | None) -> list[float] | None:
+    if parse is None:
+        return None
+    needle = _normalize(quote)
+    if not needle:
+        return None
+    for block in parse.blocks:
+        if block.bbox is not None and needle in _normalize(block.text):
+            return block.bbox
+    return None
+
+
+def flatten_answers(output: BaseModel, prefix: str = "") -> list[tuple[str, JsonValue, str | None, float]]:
+    """Walk the dynamic output model into (dotted_name, value, quote, confidence) leaves."""
+    leaves: list[tuple[str, JsonValue, str | None, float]] = []
+    for name in type(output).model_fields:
+        node = getattr(output, name)
+        dotted = f"{prefix}{name}"
+        if isinstance(node, BaseModel) and "confidence" in type(node).model_fields:
+            value = getattr(node, "value", None)
+            quote = getattr(node, "quote", None)
+            confidence = float(getattr(node, "confidence", 0.0) or 0.0)
+            leaves.append((dotted, value, quote, confidence))
+        elif isinstance(node, BaseModel):
+            leaves.extend(flatten_answers(node, prefix=f"{dotted}."))
+    return leaves
+
+
+def _format_pages(pages: list[PageText], max_characters: int) -> str:
+    parts: list[str] = []
+    used = 0
+    for page in pages:
+        snippet = page.text[: max(0, max_characters - used)]
+        parts.append(f"[Page {page.page_number}]\n{snippet}")
+        used += len(snippet)
+        if used >= max_characters:
+            break
+    return "\n\n".join(parts) if parts else "(no extractable text)"
+
+
+def pages_from_parse(parse: ParseDocumentResponse) -> list[PageText]:
+    """Rebuild per-page text from parse blocks (advanced path with no caller text)."""
+    by_page: dict[int, list[str]] = {}
+    for block in parse.blocks:
+        by_page.setdefault(block.page, []).append(block.text)
+    return [PageText(page_number=n, text="\n".join(t)) for n, t in sorted(by_page.items())]
+
+
+class ExtractFieldsAgent:
+    """One smart-model pass over the document, then code-side grounding."""
+
+    def __init__(self, runtime: AppRuntime) -> None:
+        self.runtime = runtime
+
+    async def extract(
+        self,
+        request: ExtractFieldsRequest,
+        pages: list[PageText],
+        parse: ParseDocumentResponse | None,
+    ) -> ExtractFieldsResponse:
+        output_model = build_output_model(dict(request.fields_schema))
+        provider = self.runtime.settings.chat_provider
+        agent: Agent[None, BaseModel] = Agent(
+            model=self.runtime.smart_model,
+            output_type=structured_output([output_model], chat_provider=provider),
+            system_prompt=_SYSTEM_PROMPT,
+            model_settings=self.runtime.smart_model_settings,
+            retries=output_retries(provider),
+        )
+        prompt = self._build_prompt(request, pages)
+        result = await agent.run(prompt)
+        fields = self._ground(result.output, pages, parse)
+        overall = round(min((f.confidence for f in fields), default=0.0), 4)
+        tier = DocparseTier.ADVANCED if parse is not None else DocparseTier.BASIC
+        return ExtractFieldsResponse(mode=tier, fields=fields, overall_confidence=overall)
+
+    def _build_prompt(self, request: ExtractFieldsRequest, pages: list[PageText]) -> str:
+        instructions = f"Additional instructions: {request.instructions}\n\n" if request.instructions else ""
+        return (
+            f"{instructions}"
+            f"Document file name: {request.file_name}\n"
+            f"Document content:\n{_format_pages(pages, self.runtime.settings.max_characters)}"
+        )
+
+    @staticmethod
+    def _ground(
+        output: BaseModel,
+        pages: list[PageText],
+        parse: ParseDocumentResponse | None,
+    ) -> list[ExtractedField]:
+        fields: list[ExtractedField] = []
+        for name, value, quote, model_confidence in flatten_answers(output):
+            citations: list[FieldCitation] = []
+            confidence = max(0.0, min(1.0, model_confidence))
+            if value is None:
+                confidence = 0.0
+            elif quote:
+                located = find_quote(quote, pages)
+                if located is not None:
+                    page_number, start, end = located
+                    citations.append(
+                        FieldCitation(
+                            page=page_number,
+                            bbox=_bbox_for_quote(quote, parse),
+                            quote=quote,
+                            start_offset=start,
+                            end_offset=end,
+                        )
+                    )
+                else:
+                    citations.append(FieldCitation(page=None, bbox=None, quote=quote))
+                    confidence *= UNGROUNDED_PENALTY
+            else:
+                # Terse models (local Ollama especially) often skip the quote;
+                # grounding the value itself keeps citations and a usable score.
+                value_text = _render_value(value)
+                located = find_quote(value_text, pages) if value_text else None
+                if located is not None:
+                    page_number, start, end = located
+                    citations.append(
+                        FieldCitation(
+                            page=page_number,
+                            bbox=_bbox_for_quote(value_text, parse),
+                            quote=value_text,
+                            start_offset=start,
+                            end_offset=end,
+                        )
+                    )
+                    confidence = max(confidence, VALUE_GROUNDED_FLOOR)
+                else:
+                    confidence *= UNGROUNDED_PENALTY
+            fields.append(ExtractedField(name=name, value=value, confidence=round(confidence, 4), citations=citations))
+        return fields
+
+
+def _render_value(value: JsonValue) -> str:
+    """A searchable text form of a leaf value; empty when nothing sensible exists."""
+    if value is None or isinstance(value, (dict, list)) or isinstance(value, bool):
+        return ""
+    return str(value)

engine/src/stirling/docparse/splitter.py

@@ -0,0 +1,105 @@
+"""Content-based document splitting: an LLM finds sub-document boundaries.
+
+Works entirely from caller-supplied page text (basic tier friendly); the fast
+model sees a bounded per-page preview and answers with boundary start pages,
+which are then validated in code (monotonic, in range, capped)."""
+
+from __future__ import annotations
+
+import logging
+
+from pydantic import Field
+from pydantic_ai import Agent
+
+from stirling.agents.output_mode import output_retries, structured_output
+from stirling.contracts.docparse import SmartSplitRequest, SmartSplitResponse, SplitPart
+from stirling.contracts.documents import PageText
+from stirling.models import ApiModel
+from stirling.services import AppRuntime
+
+logger = logging.getLogger(__name__)
+
+# Per-page preview budget; boundaries are recognisable from page openings.
+PAGE_PREVIEW_CHARS = 600
+
+_SYSTEM_PROMPT = (
+    "You split a multi-document file into its component documents.\n"
+    "\n"
+    "You are shown the beginning of every page. Apply the user's splitting rule and "
+    "answer with every page where a NEW component document starts.\n"
+    "Rules:\n"
+    "- Page 1 always starts the first component.\n"
+    "- Give each component a short descriptive label (e.g. 'Invoice #4821', 'Cover letter').\n"
+    "- Give your confidence 0.0-1.0 per boundary.\n"
+    "- If the rule doesn't match anything, return just the page-1 component spanning the whole file."
+)
+
+
+class _Boundary(ApiModel):
+    start_page: int = Field(ge=1, description="First page of this component document.")
+    label: str = Field(description="Short human label for the component.")
+    confidence: float = Field(ge=0.0, le=1.0)
+
+
+class _SplitOutput(ApiModel):
+    boundaries: list[_Boundary] = Field(default_factory=list)
+
+
+def _format_pages(pages: list[PageText], max_pages: int) -> str:
+    shown = pages[:max_pages]
+    parts = [f"[Page {p.page_number}] {p.text[:PAGE_PREVIEW_CHARS]}" for p in shown]
+    if len(pages) > max_pages:
+        parts.append(f"({len(pages) - max_pages} further pages omitted)")
+    return "\n\n".join(parts) if parts else "(no extractable text)"
+
+
+def validate_boundaries(output: _SplitOutput, page_count: int, max_parts: int) -> list[SplitPart]:
+    """Coerce the model's boundaries into a clean, complete partition of 1..page_count."""
+    starts: dict[int, _Boundary] = {}
+    for boundary in output.boundaries:
+        if 1 <= boundary.start_page <= page_count and boundary.start_page not in starts:
+            starts[boundary.start_page] = boundary
+    if 1 not in starts:
+        starts[1] = _Boundary(start_page=1, label="Document", confidence=1.0)
+
+    ordered = [starts[k] for k in sorted(starts)][:max_parts]
+    parts: list[SplitPart] = []
+    for i, boundary in enumerate(ordered):
+        end_page = ordered[i + 1].start_page - 1 if i + 1 < len(ordered) else page_count
+        parts.append(
+            SplitPart(
+                start_page=boundary.start_page,
+                end_page=end_page,
+                label=boundary.label.strip() or f"Part {i + 1}",
+                confidence=round(boundary.confidence, 4),
+            )
+        )
+    return parts
+
+
+class SmartSplitAgent:
+    def __init__(self, runtime: AppRuntime) -> None:
+        self.runtime = runtime
+        provider = runtime.settings.chat_provider
+        self._agent: Agent[None, _SplitOutput] = Agent(
+            model=runtime.fast_model,
+            output_type=structured_output([_SplitOutput], chat_provider=provider),
+            system_prompt=_SYSTEM_PROMPT,
+            model_settings=runtime.fast_model_settings,
+            retries=output_retries(provider),
+        )
+
+    async def split(self, request: SmartSplitRequest) -> SmartSplitResponse:
+        pages = request.pages
+        if not pages:
+            return SmartSplitResponse(parts=[])
+        page_count = max(p.page_number for p in pages)
+        prompt = (
+            f"Splitting rule: {request.rule}\n\n"
+            f"Document file name: {request.file_name}\n"
+            f"Pages:\n{_format_pages(pages, self.runtime.settings.max_pages)}"
+        )
+        result = await self._agent.run(prompt)
+        parts = validate_boundaries(result.output, page_count, request.max_parts)
+        logger.info("docparse: split %s into %d parts", request.file_name, len(parts))
+        return SmartSplitResponse(parts=parts)

engine/src/stirling/docparse/suggest_schema.py

@@ -0,0 +1,120 @@
+"""Schema suggestion: the fast model proposes extractable fields for a document.
+
+Reads a bounded window of the first pages (the fields worth extracting from a
+document type are evident from its opening) and answers with candidate fields,
+which are then validated in code: names coerced to snake_case, duplicates and
+unsupported types dropped, capped at the caller's maxFields."""
+
+from __future__ import annotations
+
+import logging
+import re
+
+from pydantic import Field
+from pydantic_ai import Agent
+
+from stirling.agents.output_mode import output_retries, structured_output
+from stirling.contracts.docparse import (
+    DocparseTier,
+    SuggestedField,
+    SuggestedFieldType,
+    SuggestSchemaRequest,
+    SuggestSchemaResponse,
+)
+from stirling.contracts.documents import PageText
+from stirling.models import ApiModel
+from stirling.services import AppRuntime
+
+logger = logging.getLogger(__name__)
+
+# First pages read; a fixed window keeps cost flat regardless of length.
+WINDOW_PAGES = 3
+# Per-page preview budget; field candidates show up near page openings.
+PAGE_PREVIEW_CHARS = 2_000
+
+_SYSTEM_PROMPT = (
+    "You design an extraction schema for a document type.\n"
+    "\n"
+    "You are shown the first pages of a document. Propose the most useful fields "
+    "a user would want extracted from documents of this type.\n"
+    "Rules:\n"
+    "- Name each field as a snake_case identifier (e.g. 'invoice_number').\n"
+    "- Type each field as one of: string, number, integer, boolean.\n"
+    "- Give each field a one-sentence description of what it holds.\n"
+    "- Propose fields for the document TYPE, not only values visible on these pages.\n"
+    "- Order fields from most to least useful."
+)
+
+
+class _SuggestedField(ApiModel):
+    # Loosely typed on purpose: bad names/types are dropped in code, not retried.
+    name: str = Field(description="snake_case identifier for the field.")
+    type: str = Field(description="One of: string, number, integer, boolean.")
+    description: str = ""
+
+
+class _SuggestOutput(ApiModel):
+    fields: list[_SuggestedField] = Field(default_factory=list)
+
+
+_IDENTIFIER = re.compile(r"[a-z][a-z0-9_]*")
+_CAMEL_BOUNDARY = re.compile(r"(?<=[a-z0-9])(?=[A-Z])")
+_NON_ALNUM = re.compile(r"[^A-Za-z0-9]+")
+_VALID_TYPES = {t.value for t in SuggestedFieldType}
+
+
+def to_snake_case(name: str) -> str:
+    """Coerce a model-proposed name into snake_case ('Invoice No.' -> 'invoice_no')."""
+    return _NON_ALNUM.sub("_", _CAMEL_BOUNDARY.sub("_", name.strip())).strip("_").lower()
+
+
+def validate_fields(output: _SuggestOutput, max_fields: int) -> list[SuggestedField]:
+    """Keep unique snake_case names with supported types, capped at ``max_fields``."""
+    kept: list[SuggestedField] = []
+    seen: set[str] = set()
+    for field in output.fields:
+        name = to_snake_case(field.name)
+        type_name = field.type.strip().lower()
+        if not _IDENTIFIER.fullmatch(name) or name in seen or type_name not in _VALID_TYPES:
+            continue
+        seen.add(name)
+        kept.append(
+            SuggestedField(name=name, type=SuggestedFieldType(type_name), description=field.description.strip())
+        )
+        if len(kept) == max_fields:
+            break
+    return kept
+
+
+def _format_pages(pages: list[PageText]) -> str:
+    shown = pages[:WINDOW_PAGES]
+    parts = [f"[Page {p.page_number}] {p.text[:PAGE_PREVIEW_CHARS]}" for p in shown]
+    if len(pages) > WINDOW_PAGES:
+        parts.append(f"({len(pages) - WINDOW_PAGES} further pages omitted)")
+    return "\n\n".join(parts) if parts else "(no extractable text)"
+
+
+class SuggestSchemaAgent:
+    def __init__(self, runtime: AppRuntime) -> None:
+        self.runtime = runtime
+        provider = runtime.settings.chat_provider
+        self._agent: Agent[None, _SuggestOutput] = Agent(
+            model=runtime.fast_model,
+            output_type=structured_output([_SuggestOutput], chat_provider=provider),
+            system_prompt=_SYSTEM_PROMPT,
+            model_settings=runtime.fast_model_settings,
+            retries=output_retries(provider),
+        )
+
+    async def suggest(
+        self, request: SuggestSchemaRequest, pages: list[PageText], tier: DocparseTier
+    ) -> SuggestSchemaResponse:
+        prompt = (
+            f"Propose up to {request.max_fields} fields.\n\n"
+            f"Document file name: {request.file_name}\n"
+            f"Document content (first pages):\n{_format_pages(pages)}"
+        )
+        result = await self._agent.run(prompt)
+        fields = validate_fields(result.output, request.max_fields)
+        logger.info("docparse: suggested %d fields for %s", len(fields), request.file_name)
+        return SuggestSchemaResponse(mode=tier, fields=fields)

engine/src/stirling/documents/__init__.py

@@ -3,11 +3,20 @@ from __future__ import annotations
 from stirling.documents.embedder import EmbeddingService
 from stirling.documents.pgvector_store import PgVectorStore
 from stirling.documents.rag_capability import RagCapability
-from stirling.documents.service import DocumentService
+from stirling.documents.service import CollectionSearchHit, DocumentService
 from stirling.documents.sqlite_vec_store import SqliteVecStore
-from stirling.documents.store import Document, DocumentStore, SearchResult, StoredPage
+from stirling.documents.store import (
+    CollectionSummary,
+    Document,
+    DocumentStore,
+    SearchResult,
+    StoredPage,
+    StoreStats,
+)
 
 __all__ = [
+    "CollectionSearchHit",
+    "CollectionSummary",
     "Document",
     "DocumentService",
     "DocumentStore",
@@ -16,5 +25,6 @@ __all__ = [
     "RagCapability",
     "SearchResult",
     "SqliteVecStore",
+    "StoreStats",
     "StoredPage",
 ]

engine/src/stirling/documents/pgvector_store.py

@@ -10,7 +10,14 @@ from pgvector.psycopg import register_vector_async
 from psycopg_pool import AsyncConnectionPool
 
 from stirling.contracts.documents import Page, PageRange
-from stirling.documents.store import Document, DocumentStore, SearchResult, StoredPage
+from stirling.documents.store import (
+    CollectionSummary,
+    Document,
+    DocumentStore,
+    SearchResult,
+    StoredPage,
+    StoreStats,
+)
 from stirling.models import OwnerId, PrincipalId
 
 _READ_PERMISSION = "read"
@@ -411,5 +418,44 @@ class PgVectorStore(DocumentStore):
                 rows = await cur.fetchall()
         return [r[0] for r in rows]
 
+    async def list_collection_summaries(self, principals: list[PrincipalId]) -> list[CollectionSummary]:
+        if not principals:
+            return []
+        await self._ensure_ready()
+        async with self._pool.connection() as conn:
+            async with conn.cursor() as cur:
+                # MIN(owner_id) mirrors _readable_owner_for's ORDER BY owner_id LIMIT 1.
+                await cur.execute(
+                    """
+                    SELECT r.collection, m.source, COUNT(d.id)
+                    FROM (
+                        SELECT collection, MIN(owner_id) AS owner_id
+                        FROM document_acl
+                        WHERE permission = %s AND principal_id = ANY(%s)
+                        GROUP BY collection
+                    ) r
+                    JOIN documents_meta m ON m.collection = r.collection AND m.owner_id = r.owner_id
+                    LEFT JOIN rag_documents d ON d.collection = r.collection AND d.owner_id = r.owner_id
+                    GROUP BY r.collection, m.source
+                    ORDER BY r.collection
+                    """,
+                    (_READ_PERMISSION, list(principals)),
+                )
+                rows = await cur.fetchall()
+        return [CollectionSummary(collection=r[0], source=r[1], chunks=int(r[2])) for r in rows]
+
+    async def stats(self) -> StoreStats:
+        await self._ensure_ready()
+        async with self._pool.connection() as conn:
+            async with conn.cursor() as cur:
+                await cur.execute("SELECT COUNT(DISTINCT collection) FROM documents_meta")
+                doc_row = await cur.fetchone()
+                await cur.execute("SELECT COUNT(*) FROM rag_documents")
+                chunk_row = await cur.fetchone()
+        return StoreStats(
+            documents=int(doc_row[0]) if doc_row else 0,
+            chunks=int(chunk_row[0]) if chunk_row else 0,
+        )
+
     async def close(self) -> None:
         await self._pool.close()

engine/src/stirling/documents/service.py

@@ -1,15 +1,32 @@
 from __future__ import annotations
 
 import logging
+from dataclasses import dataclass
 from datetime import datetime
 
 from stirling.contracts.documents import Page, PageRange, PageText
 from stirling.documents.embedder import EmbeddingService
-from stirling.documents.store import Document, DocumentStore, SearchResult, StoredPage
+from stirling.documents.store import (
+    CollectionSummary,
+    Document,
+    DocumentStore,
+    SearchResult,
+    StoredPage,
+    StoreStats,
+)
 from stirling.models import FileId, OwnerId, PrincipalId
 
 logger = logging.getLogger(__name__)
 
+
+@dataclass(frozen=True)
+class CollectionSearchHit:
+    """A search result tagged with the collection it came from."""
+
+    collection: FileId
+    result: SearchResult
+
+
 PAGE_NUMBER_METADATA_KEY = "page_number"
 CONTENT_TYPE_METADATA_KEY = "content_type"
 PAGE_TEXT_CONTENT_TYPE = "page_text"
@@ -187,6 +204,32 @@ class DocumentService:
         all_results.sort(key=lambda r: r.score, reverse=True)
         return all_results[:k]
 
+    async def search_with_collections(
+        self,
+        query: str,
+        principals: list[PrincipalId],
+        top_k: int | None = None,
+    ) -> list[CollectionSearchHit]:
+        """Cross-collection search like :meth:`search`, but every result keeps
+        the collection it came from. Restricted to what ``principals`` can read.
+        """
+        k = top_k if top_k is not None else self._default_top_k
+        query_embedding = await self._embedder.embed_query(query)
+        hits: list[CollectionSearchHit] = []
+        for col_name in await self._store.list_collections(principals):
+            try:
+                results = await self._store.search(col_name, query_embedding, k, principals)
+            except Exception:  # noqa: BLE001 - any backend error on one collection should not stop the others
+                logger.warning(
+                    "Skipping collection %s during cross-collection search",
+                    col_name,
+                    exc_info=True,
+                )
+                continue
+            hits.extend(CollectionSearchHit(collection=FileId(col_name), result=r) for r in results)
+        hits.sort(key=lambda hit: hit.result.score, reverse=True)
+        return hits[:k]
+
     async def read_pages(
         self,
         collection: FileId,
@@ -223,6 +266,10 @@ class DocumentService:
         """List collections readable by at least one of ``principals``."""
         return [FileId(name) for name in await self._store.list_collections(principals)]
 
+    async def list_documents(self, principals: list[PrincipalId]) -> list[CollectionSummary]:
+        """Per-document rollup (source, chunk count) readable by ``principals``."""
+        return await self._store.list_collection_summaries(principals)
+
     async def grant_read(
         self,
         collection: FileId,
@@ -241,6 +288,10 @@ class DocumentService:
         """Revoke a principal's access on an existing doc."""
         await self._store.revoke(collection, owner_id, principal)
 
+    async def stats(self) -> StoreStats:
+        """Deployment-wide document/chunk counts from the backing store."""
+        return await self._store.stats()
+
     async def close(self) -> None:
         """Release the underlying store's resources."""
         await self._store.close()

engine/src/stirling/documents/sqlite_vec_store.py

@@ -11,7 +11,14 @@ from pathlib import Path
 import sqlite_vec
 
 from stirling.contracts.documents import Page, PageRange
-from stirling.documents.store import Document, DocumentStore, SearchResult, StoredPage
+from stirling.documents.store import (
+    CollectionSummary,
+    Document,
+    DocumentStore,
+    SearchResult,
+    StoredPage,
+    StoreStats,
+)
 from stirling.models import OwnerId, PrincipalId
 
 _READ_PERMISSION = "read"
@@ -538,6 +545,42 @@ class SqliteVecStore(DocumentStore):
         ).fetchall()
         return [r[0] for r in rows]
 
+    async def list_collection_summaries(self, principals: list[PrincipalId]) -> list[CollectionSummary]:
+        async with self._lock:
+            return await asyncio.to_thread(self._sync_list_collection_summaries, principals)
+
+    def _sync_list_collection_summaries(self, principals: list[PrincipalId]) -> list[CollectionSummary]:
+        if not principals:
+            return []
+        placeholders = ",".join("?" * len(principals))
+        # MIN(owner_id) mirrors _readable_owner_for's ORDER BY owner_id LIMIT 1.
+        rows = self._conn.execute(
+            f"""
+            SELECT r.collection, m.source, COUNT(d.id)
+            FROM (
+                SELECT collection, MIN(owner_id) AS owner_id
+                FROM document_acl
+                WHERE permission = ? AND principal_id IN ({placeholders})
+                GROUP BY collection
+            ) r
+            JOIN documents_meta m ON m.collection = r.collection AND m.owner_id = r.owner_id
+            LEFT JOIN documents d ON d.collection = r.collection AND d.owner_id = r.owner_id
+            GROUP BY r.collection, m.source
+            ORDER BY r.collection
+            """,
+            (_READ_PERMISSION, *principals),
+        ).fetchall()
+        return [CollectionSummary(collection=r[0], source=r[1], chunks=int(r[2])) for r in rows]
+
+    async def stats(self) -> StoreStats:
+        async with self._lock:
+            return await asyncio.to_thread(self._sync_stats)
+
+    def _sync_stats(self) -> StoreStats:
+        documents = self._conn.execute("SELECT COUNT(DISTINCT collection) FROM documents_meta").fetchone()[0]
+        chunks = self._conn.execute("SELECT COUNT(*) FROM documents").fetchone()[0]
+        return StoreStats(documents=int(documents), chunks=int(chunks))
+
     async def close(self) -> None:
         async with self._lock:
             await asyncio.to_thread(self._sync_close)

engine/src/stirling/documents/store.py

@@ -34,6 +34,23 @@ class StoredPage:
     char_count: int
 
 
+@dataclass
+class StoreStats:
+    """Deployment-wide counts: distinct document ids and total vector-chunk rows."""
+
+    documents: int
+    chunks: int
+
+
+@dataclass
+class CollectionSummary:
+    """Rollup row for one readable collection: stored source label + chunk count."""
+
+    collection: str
+    source: str
+    chunks: int
+
+
 class DocumentStore(ABC):
     """Abstract interface for document storage backends.
 
@@ -148,6 +165,20 @@ class DocumentStore(ABC):
     async def list_collections(self, principals: list[PrincipalId]) -> list[str]:
         """Return collection names readable by at least one of ``principals``."""
 
+    @abstractmethod
+    async def list_collection_summaries(self, principals: list[PrincipalId]) -> list[CollectionSummary]:
+        """Per-collection rollup (source + chunk count) readable by ``principals``.
+
+        Counts cover the same owner's copy a read would resolve to, so the
+        rollup never leaks another tenant's content.
+        """
+
+    # ── deployment-wide stats (not tenant-scoped) ──────────────────────────
+
+    @abstractmethod
+    async def stats(self) -> StoreStats:
+        """Count every owner's content: distinct document ids + total chunk rows."""
+
     # ── lifecycle ──────────────────────────────────────────────────────────
 
     @abstractmethod

engine/tests/docparse/test_docxfill.py

@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+import base64
+import io
+from typing import Any
+
+import docx
+
+from stirling.contracts.docparse import FillDocxRequest
+from stirling.docparse.docxfill import fill_docx
+
+
+def _template_base64() -> str:
+    document = docx.Document()
+    document.add_paragraph("Dear {{ customer.name }},")
+    document.add_paragraph("Your total is {{ total }}.")
+    document.add_paragraph("Unknown: {{ nowhere.field }}")
+    table = document.add_table(rows=2, cols=2)
+    table.rows[0].cells[0].text = "Item"
+    table.rows[0].cells[1].text = "Price"
+    table.rows[1].cells[0].text = "{{#items.name}}"
+    table.rows[1].cells[1].text = "{{#items.price}}"
+    buffer = io.BytesIO()
+    document.save(buffer)
+    return base64.b64encode(buffer.getvalue()).decode("ascii")
+
+
+def _load(response_base64: str) -> Any:
+    return docx.Document(io.BytesIO(base64.b64decode(response_base64)))
+
+
+def test_fills_scalars_tables_and_reports_missing() -> None:
+    request = FillDocxRequest(
+        template_base64=_template_base64(),
+        data={
+            "customer": {"name": "ACME GmbH"},
+            "total": 12.5,
+            "items": [
+                {"name": "Widget", "price": "2.00"},
+                {"name": "Gadget", "price": "10.50"},
+            ],
+        },
+    )
+    response = fill_docx(request)
+    filled = _load(response.docx_base64)
+
+    paragraphs = [p.text for p in filled.paragraphs]
+    assert "Dear ACME GmbH," in paragraphs
+    assert "Your total is 12.5." in paragraphs
+    # Unresolved placeholders stay put and are reported.
+    assert any("{{ nowhere.field }}" in p for p in paragraphs)
+    assert response.missing == ["nowhere.field"]
+
+    table = filled.tables[0]
+    rendered_rows = [[cell.text for cell in row.cells] for row in table.rows]
+    assert ["Widget", "2.00"] in rendered_rows
+    assert ["Gadget", "10.50"] in rendered_rows
+    # The template row is gone.
+    assert all("{{#" not in cell for row in rendered_rows for cell in row)
+    assert response.replaced >= 6
+
+
+def test_empty_items_removes_template_row() -> None:
+    request = FillDocxRequest(
+        template_base64=_template_base64(),
+        data={"customer": {"name": "X"}, "total": 1, "items": []},
+    )
+    response = fill_docx(request)
+    filled = _load(response.docx_base64)
+    assert len(filled.tables[0].rows) == 1  # only the header remains

engine/tests/docparse/test_grounding.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+from pydantic import BaseModel
+
+from stirling.contracts import PageText
+from stirling.docparse.extractor import (
+    VALUE_GROUNDED_FLOOR,
+    ExtractFieldsAgent,
+    build_output_model,
+    find_quote,
+)
+from stirling.docparse.splitter import _Boundary, _SplitOutput, validate_boundaries
+
+
+def _pages() -> list[PageText]:
+    return [
+        PageText(page_number=1, text="Invoice INV-123\nTotal due:   1,240.00 EUR"),
+        PageText(page_number=2, text="Payment terms\nNet 30 days from receipt."),
+    ]
+
+
+def test_find_quote_exact() -> None:
+    located = find_quote("Invoice INV-123", _pages())
+    assert located is not None
+    page, start, end = located
+    assert page == 1
+    assert start == 0
+
+
+def test_find_quote_is_whitespace_insensitive() -> None:
+    located = find_quote("Total due: 1,240.00 EUR", _pages())
+    assert located is not None
+    assert located[0] == 1
+
+
+def test_find_quote_is_case_insensitive_and_crosses_pages() -> None:
+    located = find_quote("net 30 DAYS", _pages())
+    assert located is not None
+    assert located[0] == 2
+
+
+def test_find_quote_missing_returns_none() -> None:
+    assert find_quote("does not appear", _pages()) is None
+    assert find_quote("   ", _pages()) is None
+
+
+def test_validate_boundaries_partitions_cleanly() -> None:
+    output = _SplitOutput(
+        boundaries=[
+            _Boundary(start_page=4, label="Invoice B", confidence=0.8),
+            _Boundary(start_page=1, label="Invoice A", confidence=0.9),
+            _Boundary(start_page=4, label="dup", confidence=0.1),
+            _Boundary(start_page=99, label="out of range", confidence=0.5),
+        ]
+    )
+    parts = validate_boundaries(output, page_count=6, max_parts=10)
+    assert [(p.start_page, p.end_page) for p in parts] == [(1, 3), (4, 6)]
+    assert parts[0].label == "Invoice A"
+
+
+def test_validate_boundaries_inserts_page_one() -> None:
+    output = _SplitOutput(boundaries=[_Boundary(start_page=3, label="Part", confidence=0.7)])
+    parts = validate_boundaries(output, page_count=5, max_parts=10)
+    assert parts[0].start_page == 1
+    assert parts[1].start_page == 3
+    assert parts[-1].end_page == 5
+
+
+def test_validate_boundaries_empty_output_spans_whole_file() -> None:
+    parts = validate_boundaries(_SplitOutput(), page_count=7, max_parts=10)
+    assert [(p.start_page, p.end_page) for p in parts] == [(1, 7)]
+
+
+def _answers(quote: str | None, confidence: float) -> BaseModel:
+    model = build_output_model({"type": "object", "properties": {"invoice_number": {"type": "string"}}})
+    return model.model_validate({"invoiceNumber": {"value": "INV-123", "quote": quote, "confidence": confidence}})
+
+
+def test_ground_falls_back_to_value_when_quote_missing() -> None:
+    # Terse local models return the value but no quote; the value itself grounds.
+    pages = [PageText(page_number=1, text="Invoice INV-123 issued today.")]
+    fields = ExtractFieldsAgent._ground(_answers(quote=None, confidence=0.0), pages, None)
+    assert fields[0].citations and fields[0].citations[0].page == 1
+    assert fields[0].citations[0].quote == "INV-123"
+    assert fields[0].confidence >= VALUE_GROUNDED_FLOOR
+
+
+def test_ground_penalises_when_nothing_grounds() -> None:
+    pages = [PageText(page_number=1, text="completely unrelated text")]
+    fields = ExtractFieldsAgent._ground(_answers(quote=None, confidence=0.9), pages, None)
+    assert not fields[0].citations
+    assert fields[0].confidence < 0.9

engine/tests/docparse/test_schema_model.py

@@ -0,0 +1,79 @@
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+from stirling.docparse.extractor import MAX_SCHEMA_DEPTH, SchemaError, build_output_model, flatten_answers
+
+
+def _schema(properties: dict[str, Any]) -> dict[str, Any]:
+    return {"type": "object", "properties": properties}
+
+
+def test_builds_scalar_fields() -> None:
+    model = build_output_model(
+        _schema(
+            {
+                "invoice_number": {"type": "string", "description": "The invoice id"},
+                "total": {"type": "number"},
+                "line_count": {"type": "integer"},
+                "paid": {"type": "boolean"},
+            }
+        )
+    )
+    instance = model.model_validate(
+        {
+            "invoiceNumber": {"value": "INV-1", "quote": "Invoice INV-1", "confidence": 0.9},
+            "total": {"value": 12.5, "quote": None, "confidence": 0.8},
+            "lineCount": {"value": 3, "quote": None, "confidence": 0.7},
+            "paid": {"value": True, "quote": None, "confidence": 0.6},
+        }
+    )
+    leaves = dict((name, value) for name, value, _quote, _conf in flatten_answers(instance))
+    assert leaves == {"invoice_number": "INV-1", "total": 12.5, "line_count": 3, "paid": True}
+
+
+def test_nested_objects_flatten_to_dotted_names() -> None:
+    model = build_output_model(_schema({"vendor": {"type": "object", "properties": {"name": {"type": "string"}}}}))
+    instance = model.model_validate({"vendor": {"name": {"value": "ACME", "quote": None, "confidence": 0.5}}})
+    names = [name for name, _v, _q, _c in flatten_answers(instance)]
+    assert names == ["vendor.name"]
+
+
+def test_arrays_of_scalars() -> None:
+    model = build_output_model(_schema({"tags": {"type": "array", "items": {"type": "string"}}}))
+    instance = model.model_validate({"tags": {"value": ["a", "b"], "quote": None, "confidence": 1.0}})
+    leaves = flatten_answers(instance)
+    assert leaves[0][1] == ["a", "b"]
+
+
+def test_enum_lands_in_description_not_type() -> None:
+    model = build_output_model(_schema({"currency": {"type": "string", "enum": ["EUR", "USD"]}}))
+    answer_model = model.model_fields["currency"].annotation
+    description = answer_model.model_fields["value"].description  # type: ignore[union-attr]
+    assert "EUR" in description and "USD" in description
+
+
+@pytest.mark.parametrize(
+    "schema",
+    [
+        {"type": "object", "properties": {}},
+        {"type": "object"},
+        {"type": "object", "properties": {"bad-name": {"type": "string"}}},
+        {"type": "object", "properties": {"x": {"type": "date"}}},
+        {"type": "object", "properties": {"x": {"type": "array"}}},
+        {"type": "object", "properties": {"x": {"type": "array", "items": {"type": "object"}}}},
+    ],
+)
+def test_rejects_unsupported_schemas(schema: dict[str, Any]) -> None:
+    with pytest.raises(SchemaError):
+        build_output_model(schema)
+
+
+def test_rejects_over_deep_nesting() -> None:
+    schema: dict = {"type": "string"}
+    for _ in range(MAX_SCHEMA_DEPTH + 2):
+        schema = {"type": "object", "properties": {"child": schema}}
+    with pytest.raises(SchemaError):
+        build_output_model(schema)

engine/tests/docparse/test_suggest_schema.py

@@ -0,0 +1,129 @@
+from __future__ import annotations
+
+import base64
+from collections.abc import Iterator
+
+import pytest
+from fastapi.testclient import TestClient
+
+from stirling.api import app
+from stirling.api.dependencies import get_suggest_schema_agent
+from stirling.api.routes import docparse as docparse_routes
+from stirling.contracts.docparse import (
+    DocparseCapabilities,
+    DocparseTier,
+    SuggestedField,
+    SuggestedFieldType,
+    SuggestSchemaRequest,
+    SuggestSchemaResponse,
+)
+from stirling.contracts.documents import PageText
+from stirling.docparse.suggest_schema import _SuggestedField, _SuggestOutput, to_snake_case, validate_fields
+
+
+def _force_addon(monkeypatch: pytest.MonkeyPatch, installed: bool) -> None:
+    caps = DocparseCapabilities(advanced_installed=installed, models_available=installed)
+    monkeypatch.setattr(docparse_routes, "probe_capabilities", lambda _home, refresh=False: caps)
+
+
+class StubSuggestAgent:
+    def __init__(self) -> None:
+        self.seen_pages: list[PageText] | None = None
+
+    async def suggest(
+        self, _request: SuggestSchemaRequest, pages: list[PageText], tier: DocparseTier
+    ) -> SuggestSchemaResponse:
+        self.seen_pages = pages
+        return SuggestSchemaResponse(
+            mode=tier,
+            fields=[SuggestedField(name="invoice_number", type=SuggestedFieldType.STRING, description="The number.")],
+        )
+
+
+@pytest.fixture
+def stub_agent() -> StubSuggestAgent:
+    return StubSuggestAgent()
+
+
+@pytest.fixture
+def client(stub_agent: StubSuggestAgent) -> Iterator[TestClient]:
+    app.dependency_overrides[get_suggest_schema_agent] = lambda: stub_agent
+    try:
+        yield TestClient(app)
+    finally:
+        app.dependency_overrides.pop(get_suggest_schema_agent, None)
+
+
+# ── validation (model proposes, code decides) ───────────────────────────
+
+
+def _output(*fields: tuple[str, str]) -> _SuggestOutput:
+    return _SuggestOutput(fields=[_SuggestedField(name=n, type=t, description="d") for n, t in fields])
+
+
+def test_names_are_coerced_to_snake_case() -> None:
+    fields = validate_fields(_output(("Invoice Number", "string"), ("dueDate", "string")), max_fields=8)
+    assert [f.name for f in fields] == ["invoice_number", "due_date"]
+
+
+def test_invalid_types_are_dropped() -> None:
+    fields = validate_fields(_output(("total", "money"), ("count", "integer"), ("date", "datetime")), max_fields=8)
+    assert [f.name for f in fields] == ["count"]
+    assert fields[0].type is SuggestedFieldType.INTEGER
+
+
+def test_duplicate_names_collapse_to_first() -> None:
+    fields = validate_fields(_output(("total", "number"), ("Total", "string"), ("total", "integer")), max_fields=8)
+    assert len(fields) == 1
+    assert fields[0].type is SuggestedFieldType.NUMBER
+
+
+def test_result_is_capped_at_max_fields() -> None:
+    fields = validate_fields(_output(*[(f"field_{i}", "string") for i in range(10)]), max_fields=3)
+    assert [f.name for f in fields] == ["field_0", "field_1", "field_2"]
+
+
+def test_names_that_cannot_become_identifiers_are_dropped() -> None:
+    fields = validate_fields(_output(("123abc", "string"), ("!!!", "string"), ("ok_name", "string")), max_fields=8)
+    assert [f.name for f in fields] == ["ok_name"]
+
+
+def test_snake_case_coercion_examples() -> None:
+    assert to_snake_case("Invoice No.") == "invoice_no"
+    assert to_snake_case("invoiceNumber") == "invoice_number"
+    assert to_snake_case("TotalUSD") == "total_usd"
+    assert to_snake_case("  already_snake ") == "already_snake"
+
+
+# ── route ───────────────────────────────────────────────────────────────
+
+
+def test_suggest_schema_basic_tier_from_pages(client: TestClient) -> None:
+    response = client.post(
+        "/api/v1/docparse/suggest-schema",
+        json={"fileName": "invoice.pdf", "pages": [{"pageNumber": 1, "text": "Invoice INV-1"}]},
+    )
+    assert response.status_code == 200
+    body = response.json()
+    assert body["mode"] == "basic"
+    assert body["fields"][0] == {"name": "invoice_number", "type": "string", "description": "The number."}
+
+
+def test_suggest_schema_without_pages_or_content_is_422(client: TestClient) -> None:
+    response = client.post("/api/v1/docparse/suggest-schema", json={"fileName": "x.pdf"})
+    assert response.status_code == 422
+
+
+def test_suggest_schema_content_without_addon_is_422(client: TestClient, monkeypatch: pytest.MonkeyPatch) -> None:
+    _force_addon(monkeypatch, installed=False)
+    payload = {"fileName": "scan.pdf", "contentBase64": base64.b64encode(b"%PDF-1.4").decode()}
+    response = client.post("/api/v1/docparse/suggest-schema", json=payload)
+    assert response.status_code == 422
+
+
+def test_suggest_schema_rejects_max_fields_above_cap(client: TestClient) -> None:
+    response = client.post(
+        "/api/v1/docparse/suggest-schema",
+        json={"fileName": "x.pdf", "pages": [{"pageNumber": 1, "text": "t"}], "maxFields": 21},
+    )
+    assert response.status_code == 422

engine/tests/test_documents.py

@@ -7,7 +7,7 @@ from stirling.documents.chunker import chunk_text
 from stirling.documents.rag_capability import RagCapability
 from stirling.documents.service import DocumentService
 from stirling.documents.sqlite_vec_store import SqliteVecStore
-from stirling.documents.store import Document, SearchResult
+from stirling.documents.store import CollectionSummary, Document, SearchResult
 from stirling.models import FileId, OwnerId, PrincipalId
 
 # Personal-doc tests reuse the same opaque string in all three roles — keeps the
@@ -178,6 +178,27 @@ class TestSqliteVecStore:
         assert await store.list_collections(OWNER_PRINCIPALS) == []
         assert await store.list_collections(OTHER_OWNER_PRINCIPALS) == ["c.pdf"]
 
+    @pytest.mark.anyio
+    async def test_stats_count_distinct_document_ids_and_chunk_rows(self) -> None:
+        """Stats span every owner; a document id shared by two owners counts once."""
+        store = SqliteVecStore.ephemeral()
+        empty = await store.stats()
+        assert (empty.documents, empty.chunks) == (0, 0)
+
+        for owner, principals, name, texts in (
+            (OWNER, OWNER_PRINCIPALS, "doc-a", ["one", "two"]),
+            (OWNER, OWNER_PRINCIPALS, "doc-b", ["three"]),
+            (OTHER_OWNER, OTHER_OWNER_PRINCIPALS, "doc-b", ["four"]),
+        ):
+            await store.ensure_collection(name, f"{name}.pdf", owner, None)
+            await store.grant_read(name, owner, principals)
+            docs = [Document(id=str(i), text=t, metadata={}) for i, t in enumerate(texts)]
+            await store.add_documents(name, docs, [[1.0, 0.0]] * len(docs), owner)
+
+        stats = await store.stats()
+        assert stats.documents == 2
+        assert stats.chunks == 4
+
     @pytest.mark.anyio
     async def test_reap_expired_drops_collections_past_expires_at(self) -> None:
         """TTL backstop: rows with ``expires_at`` in the past go away on reap."""
@@ -239,6 +260,47 @@ class TestSqliteVecStore:
         # Owner still can.
         assert await store.has_collection("doc", OWNER_PRINCIPALS) is True
 
+    @pytest.mark.anyio
+    async def test_list_collection_summaries_rolls_up_readable_collections(self) -> None:
+        """Rollup: one row per readable collection with its source and chunk count."""
+        store = SqliteVecStore.ephemeral()
+        await store.ensure_collection("doc-a", "a.pdf", OWNER, None)
+        await store.grant_read("doc-a", OWNER, OWNER_PRINCIPALS)
+        docs = [Document(id="1", text="one", metadata={}), Document(id="2", text="two", metadata={})]
+        await store.add_documents("doc-a", docs, [[1.0, 0.0], [0.0, 1.0]], OWNER)
+        # Collection with no vector chunks yet: still listed, zero count.
+        await store.ensure_collection("doc-b", "b.pdf", OWNER, None)
+        await store.grant_read("doc-b", OWNER, OWNER_PRINCIPALS)
+
+        summaries = await store.list_collection_summaries(OWNER_PRINCIPALS)
+        assert summaries == [
+            CollectionSummary(collection="doc-a", source="a.pdf", chunks=2),
+            CollectionSummary(collection="doc-b", source="b.pdf", chunks=0),
+        ]
+
+    @pytest.mark.anyio
+    async def test_list_collection_summaries_scoped_to_principals(self) -> None:
+        """One principal's rollup never lists, or counts, another owner's copy."""
+        store = SqliteVecStore.ephemeral()
+        await store.ensure_collection("shared-id", "alice.pdf", OWNER, None)
+        await store.grant_read("shared-id", OWNER, OWNER_PRINCIPALS)
+        await store.add_documents("shared-id", [Document(id="1", text="alice", metadata={})], [[1.0, 0.0]], OWNER)
+        await store.ensure_collection("shared-id", "bob.pdf", OTHER_OWNER, None)
+        await store.grant_read("shared-id", OTHER_OWNER, OTHER_OWNER_PRINCIPALS)
+        bob_docs = [Document(id="1", text="bob", metadata={}), Document(id="2", text="bob2", metadata={})]
+        await store.add_documents("shared-id", bob_docs, [[1.0, 0.0], [0.0, 1.0]], OTHER_OWNER)
+        await store.ensure_collection("bob-only", "bob-only.pdf", OTHER_OWNER, None)
+        await store.grant_read("bob-only", OTHER_OWNER, OTHER_OWNER_PRINCIPALS)
+
+        assert await store.list_collection_summaries(OWNER_PRINCIPALS) == [
+            CollectionSummary(collection="shared-id", source="alice.pdf", chunks=1)
+        ]
+        assert await store.list_collection_summaries(OTHER_OWNER_PRINCIPALS) == [
+            CollectionSummary(collection="bob-only", source="bob-only.pdf", chunks=0),
+            CollectionSummary(collection="shared-id", source="bob.pdf", chunks=2),
+        ]
+        assert await store.list_collection_summaries([]) == []
+
 
 # DocumentService (with stub embedder)
 
@@ -398,6 +460,31 @@ class TestDocumentService:
         multi_results = await documents.search("deploy", principals=[hr_group, eng_group])
         assert len(multi_results) > 0
 
+    @pytest.mark.anyio
+    async def test_search_with_collections_tags_results_and_respects_acl(self, documents: DocumentService) -> None:
+        """Collection-tagged search only reaches collections the caller can read."""
+        await documents.ingest(
+            FileId("col-a"),
+            _pages("Alpha content."),
+            source="a.pdf",
+            owner_id=OWNER,
+            read_principals=OWNER_PRINCIPALS,
+            expires_at=None,
+        )
+        await documents.ingest(
+            FileId("col-b"),
+            _pages("Beta content."),
+            source="b.pdf",
+            owner_id=OTHER_OWNER,
+            read_principals=OTHER_OWNER_PRINCIPALS,
+            expires_at=None,
+        )
+
+        hits = await documents.search_with_collections("content", principals=OWNER_PRINCIPALS)
+        assert hits
+        assert {hit.collection for hit in hits} == {"col-a"}
+        assert all(hit.result.document.text for hit in hits)
+
     @pytest.mark.anyio
     async def test_delete_collection(self, documents: DocumentService) -> None:
         await documents.ingest(

engine/tests/test_documents_routes.py

@@ -6,9 +6,10 @@ import pytest
 from fastapi.testclient import TestClient
 
 from stirling.api import app
-from stirling.api.dependencies import get_document_service
+from stirling.api.dependencies import get_document_service, get_knowledge_ask_agent
+from stirling.contracts import AskDocumentsRequest, AskDocumentsResponse, DocumentPassage
 from stirling.documents import Document, DocumentService, SqliteVecStore
-from stirling.models import FileId, PrincipalId, UserId
+from stirling.models import FileId, OwnerId, PrincipalId, UserId
 
 USER = UserId("test-user")
 USER_PRINCIPALS = [PrincipalId("test-user")]
@@ -348,6 +349,274 @@ def test_purge_by_owner_rejects_missing_user_header(client: TestClient) -> None:
     assert response.status_code == 401
 
 
+# ── GET /documents/list ─────────────────────────────────────────────────
+
+
+def _ingest(client: TestClient, document_id: str, source: str, texts: list[str], owner: str) -> None:
+    client.post(
+        "/api/v1/documents",
+        json={
+            "documentId": document_id,
+            "source": source,
+            "pageText": [{"pageNumber": i, "text": t} for i, t in enumerate(texts, 1)],
+            "ownerId": owner,
+            "readPrincipals": [owner],
+            "expiresAt": None,
+        },
+        headers={"X-User-Id": owner},
+    )
+
+
+def test_list_documents_returns_caller_rollup(client: TestClient) -> None:
+    _ingest(client, "list-a", "a.pdf", ["Page one text.", "Page two text."], USER)
+    _ingest(client, "list-b", "b.pdf", ["Only page."], USER)
+
+    response = client.get("/api/v1/documents/list", headers=HEADERS)
+    assert response.status_code == 200
+    documents = response.json()["documents"]
+    assert [d["documentId"] for d in documents] == ["list-a", "list-b"]
+    by_id = {d["documentId"]: d for d in documents}
+    assert by_id["list-a"]["source"] == "a.pdf"
+    assert by_id["list-a"]["chunks"] >= 2
+    assert by_id["list-b"]["source"] == "b.pdf"
+    assert by_id["list-b"]["chunks"] >= 1
+
+
+def test_list_documents_empty_for_new_user(client: TestClient) -> None:
+    response = client.get("/api/v1/documents/list", headers=HEADERS)
+    assert response.status_code == 200
+    assert response.json() == {"documents": []}
+
+
+def test_list_documents_hides_other_users_documents(client: TestClient) -> None:
+    """User A must never see user B's documents in the rollup."""
+    _ingest(client, "alice-doc", "alice.pdf", ["alice content"], "alice")
+    _ingest(client, "bob-doc", "bob.pdf", ["bob content"], "bob")
+
+    alice_docs = client.get("/api/v1/documents/list", headers={"X-User-Id": "alice"}).json()["documents"]
+    bob_docs = client.get("/api/v1/documents/list", headers={"X-User-Id": "bob"}).json()["documents"]
+    assert [d["documentId"] for d in alice_docs] == ["alice-doc"]
+    assert [d["documentId"] for d in bob_docs] == ["bob-doc"]
+
+
+def test_list_documents_rejects_missing_user_header(client: TestClient) -> None:
+    assert client.get("/api/v1/documents/list").status_code == 401
+
+
+# ── POST /documents/search ──────────────────────────────────────────────
+
+
+def test_search_documents_maps_page_text_chunks(client: TestClient) -> None:
+    """Plain ingested chunks only carry page_number: both bounds map to it and
+    the ":page:N" suffix is stripped off the source."""
+    client.post(
+        "/api/v1/documents",
+        json={
+            "documentId": "report",
+            "source": "report.pdf",
+            "pageText": [{"pageNumber": 3, "text": "The launch is planned for October."}],
+            "ownerId": USER,
+            "readPrincipals": [USER],
+            "expiresAt": None,
+        },
+        headers=HEADERS,
+    )
+
+    response = client.post("/api/v1/documents/search", json={"query": "launch", "topK": 5}, headers=HEADERS)
+    assert response.status_code == 200
+    passages = response.json()["passages"]
+    assert len(passages) >= 1
+    passage = passages[0]
+    assert passage["documentId"] == "report"
+    assert passage["pageStart"] == 3
+    assert passage["pageEnd"] == 3
+    assert passage["headingPath"] == []
+    assert passage["source"] == "report.pdf"
+    assert "launch" in passage["text"]
+    assert isinstance(passage["score"], float)
+
+
+@pytest.mark.anyio
+async def test_search_documents_maps_docparse_chunk_metadata(client: TestClient, service: DocumentService) -> None:
+    """Docparse chunks carry page bounds + heading path; they map straight onto the wire."""
+    await service.ingest_prepared(
+        collection=FileId("dp-doc"),
+        chunks=[
+            (
+                "Revenue grew 12% in Q2.",
+                {
+                    "content_type": "docparse_chunk",
+                    "page_start": "2",
+                    "page_end": "3",
+                    "heading_path": "Report > Finance",
+                },
+            )
+        ],
+        source="q2.pdf",
+        owner_id=OwnerId(USER),
+        read_principals=USER_PRINCIPALS,
+        expires_at=None,
+    )
+
+    response = client.post("/api/v1/documents/search", json={"query": "revenue"}, headers=HEADERS)
+    assert response.status_code == 200
+    passage = response.json()["passages"][0]
+    assert passage["documentId"] == "dp-doc"
+    assert passage["pageStart"] == 2
+    assert passage["pageEnd"] == 3
+    assert passage["headingPath"] == ["Report", "Finance"]
+    assert passage["source"] == "q2.pdf"
+
+
+def test_search_documents_cannot_see_other_users_documents(client: TestClient) -> None:
+    """User B searching for user A's content must get nothing back."""
+    _ingest(client, "alice-doc", "alice.pdf", ["The secret launch code is October."], "alice")
+
+    bob = client.post("/api/v1/documents/search", json={"query": "secret launch"}, headers={"X-User-Id": "bob"})
+    assert bob.status_code == 200
+    assert bob.json()["passages"] == []
+
+    alice = client.post("/api/v1/documents/search", json={"query": "secret launch"}, headers={"X-User-Id": "alice"})
+    assert alice.json()["passages"] != []
+
+
+def test_search_documents_rejects_empty_query(client: TestClient) -> None:
+    response = client.post("/api/v1/documents/search", json={"query": ""}, headers=HEADERS)
+    assert response.status_code == 422
+
+
+def test_search_documents_rejects_top_k_above_cap(client: TestClient) -> None:
+    response = client.post("/api/v1/documents/search", json={"query": "x", "topK": 51}, headers=HEADERS)
+    assert response.status_code == 422
+
+
+def test_search_documents_rejects_missing_user_header(client: TestClient) -> None:
+    assert client.post("/api/v1/documents/search", json={"query": "x"}).status_code == 401
+
+
+# ── POST /documents/ask ─────────────────────────────────────────────────
+
+
+class StubKnowledgeAskAgent:
+    """Stands in for KnowledgeAskAgent so route tests don't call a model."""
+
+    def __init__(self, response: AskDocumentsResponse) -> None:
+        self._response = response
+        self.calls: list[tuple[AskDocumentsRequest, list[PrincipalId]]] = []
+
+    async def ask(self, request: AskDocumentsRequest, principals: list[PrincipalId]) -> AskDocumentsResponse:
+        self.calls.append((request, principals))
+        return self._response
+
+
+@pytest.fixture
+def ask_agent() -> StubKnowledgeAskAgent:
+    return StubKnowledgeAskAgent(
+        AskDocumentsResponse(
+            answer="Revenue grew 12% (q2.pdf p.2).",
+            passages=[
+                DocumentPassage(
+                    document_id=FileId("dp-doc"),
+                    text="Revenue grew 12% in Q2.",
+                    score=0.91,
+                    page_start=2,
+                    page_end=3,
+                    heading_path=["Report", "Finance"],
+                    source="q2.pdf",
+                )
+            ],
+        )
+    )
+
+
+@pytest.fixture
+def ask_client(client: TestClient, ask_agent: StubKnowledgeAskAgent) -> Iterator[TestClient]:
+    app.dependency_overrides[get_knowledge_ask_agent] = lambda: ask_agent
+    try:
+        yield client
+    finally:
+        app.dependency_overrides.pop(get_knowledge_ask_agent, None)
+
+
+def test_ask_documents_returns_answer_and_passages(ask_client: TestClient) -> None:
+    response = ask_client.post("/api/v1/documents/ask", json={"question": "How did revenue do?"}, headers=HEADERS)
+    assert response.status_code == 200
+    body = response.json()
+    assert body["answer"] == "Revenue grew 12% (q2.pdf p.2)."
+    assert body["passages"] == [
+        {
+            "documentId": "dp-doc",
+            "text": "Revenue grew 12% in Q2.",
+            "score": 0.91,
+            "pageStart": 2,
+            "pageEnd": 3,
+            "headingPath": ["Report", "Finance"],
+            "source": "q2.pdf",
+        }
+    ]
+
+
+def test_ask_documents_scopes_to_calling_user(ask_client: TestClient, ask_agent: StubKnowledgeAskAgent) -> None:
+    """The route hands the agent exactly the caller's principal set."""
+    ask_client.post("/api/v1/documents/ask", json={"question": "anything"}, headers=HEADERS)
+    request, principals = ask_agent.calls[0]
+    assert principals == [PrincipalId(USER)]
+    assert request.top_k == 8
+
+
+def test_ask_documents_rejects_empty_question(ask_client: TestClient) -> None:
+    response = ask_client.post("/api/v1/documents/ask", json={"question": ""}, headers=HEADERS)
+    assert response.status_code == 422
+
+
+def test_ask_documents_rejects_top_k_above_cap(ask_client: TestClient) -> None:
+    response = ask_client.post("/api/v1/documents/ask", json={"question": "x", "topK": 21}, headers=HEADERS)
+    assert response.status_code == 422
+
+
+def test_ask_documents_rejects_missing_user_header(ask_client: TestClient) -> None:
+    assert ask_client.post("/api/v1/documents/ask", json={"question": "x"}).status_code == 401
+
+
+# ── GET /documents/stats ────────────────────────────────────────────────
+
+
+def test_stats_on_empty_store_reports_zero(client: TestClient) -> None:
+    response = client.get("/api/v1/documents/stats", headers=HEADERS)
+    assert response.status_code == 200
+    body = response.json()
+    assert body["documents"] == 0
+    assert body["chunks"] == 0
+    assert body["backend"] in ("sqlite", "pgvector")
+    assert body["embeddingModel"]
+
+
+def test_stats_counts_seeded_documents_across_owners(client: TestClient) -> None:
+    """Stats are deployment-wide: both owners' content is counted."""
+    for owner, doc in (("alice", "doc-a"), ("bob", "doc-b")):
+        client.post(
+            "/api/v1/documents",
+            json={
+                "documentId": doc,
+                "source": f"{doc}.pdf",
+                "pageText": [{"pageNumber": 1, "text": "Some content for the stats endpoint."}],
+                "ownerId": owner,
+                "readPrincipals": [owner],
+                "expiresAt": None,
+            },
+            headers={"X-User-Id": owner},
+        )
+    response = client.get("/api/v1/documents/stats", headers=HEADERS)
+    assert response.status_code == 200
+    body = response.json()
+    assert body["documents"] == 2
+    assert body["chunks"] >= 2
+
+
+def test_stats_rejects_missing_user_header(client: TestClient) -> None:
+    assert client.get("/api/v1/documents/stats").status_code == 401
+
+
 def test_delete_document_only_affects_calling_user(client: TestClient) -> None:
     """Two users with the same document id: one user's delete must not remove the other's."""
     alice_body = {

engine/tests/test_knowledge_ask.py

@@ -0,0 +1,91 @@
+from __future__ import annotations
+
+import pytest
+
+from stirling.agents.knowledge_ask import KnowledgeAskAgent, format_passages, passage_from_hit
+from stirling.config import AppSettings
+from stirling.contracts import AskDocumentsRequest, DocumentPassage
+from stirling.documents import CollectionSearchHit, Document, DocumentService, SearchResult, SqliteVecStore
+from stirling.models import FileId, PrincipalId
+from stirling.services import build_runtime
+
+PRINCIPALS = [PrincipalId("test-user")]
+
+
+def _hit(metadata: dict[str, str], text: str = "chunk text", score: float = 0.8) -> CollectionSearchHit:
+    return CollectionSearchHit(
+        collection=FileId("doc-1"),
+        result=SearchResult(document=Document(id="c1", text=text, metadata=metadata), score=score),
+    )
+
+
+# ── passage_from_hit ────────────────────────────────────────────────────
+
+
+def test_passage_from_hit_maps_docparse_metadata() -> None:
+    passage = passage_from_hit(
+        _hit({"source": "q2.pdf", "page_start": "2", "page_end": "3", "heading_path": "Report > Finance"})
+    )
+    assert passage.document_id == "doc-1"
+    assert passage.page_start == 2
+    assert passage.page_end == 3
+    assert passage.heading_path == ["Report", "Finance"]
+    assert passage.source == "q2.pdf"
+
+
+def test_passage_from_hit_falls_back_to_page_number() -> None:
+    """Plain page-text chunks: page_number fills both bounds, source drops the page suffix."""
+    passage = passage_from_hit(_hit({"source": "report.pdf:page:4", "page_number": "4"}))
+    assert passage.page_start == 4
+    assert passage.page_end == 4
+    assert passage.heading_path == []
+    assert passage.source == "report.pdf"
+
+
+def test_passage_from_hit_tolerates_missing_and_bad_metadata() -> None:
+    passage = passage_from_hit(_hit({"page_start": "not-a-number"}))
+    assert passage.page_start is None
+    assert passage.page_end is None
+    assert passage.heading_path == []
+    assert passage.source is None
+
+
+# ── format_passages ─────────────────────────────────────────────────────
+
+
+def test_format_passages_includes_citation_handles() -> None:
+    passages = [
+        DocumentPassage(document_id=FileId("d1"), text="Alpha.", score=0.9, page_start=2, page_end=3, source="a.pdf"),
+        DocumentPassage(document_id=FileId("d2"), text="Beta.", score=0.5, page_start=7, page_end=7, source="b.pdf"),
+        DocumentPassage(document_id=FileId("d3"), text="Gamma.", score=0.4),
+    ]
+    rendered = format_passages(passages)
+    assert "[Passage 1 | a.pdf p.2-3]\nAlpha." in rendered
+    assert "[Passage 2 | b.pdf p.7]\nBeta." in rendered
+    # No source or pages: fall back to the document id alone.
+    assert "[Passage 3 | d3]\nGamma." in rendered
+
+
+# ── KnowledgeAskAgent ───────────────────────────────────────────────────
+
+
+class _StubEmbedder:
+    """Deterministic embeddings so the agent test needs no provider."""
+
+    async def embed_query(self, text: str) -> list[float]:
+        return [1.0, 0.0]
+
+    async def embed_documents(self, texts: list[str]) -> list[list[float]]:
+        return [[1.0, 0.0] for _ in texts]
+
+
+@pytest.mark.anyio
+async def test_ask_answers_plainly_when_nothing_retrieved(app_settings: AppSettings) -> None:
+    """Empty retrieval short-circuits: no model call, honest not-found answer."""
+    documents = DocumentService(embedder=_StubEmbedder(), store=SqliteVecStore.ephemeral(), default_top_k=3)  # type: ignore[arg-type]
+    runtime = build_runtime(app_settings, documents=documents)
+    agent = KnowledgeAskAgent(runtime)
+
+    response = await agent.ask(AskDocumentsRequest(question="What is the launch date?"), principals=PRINCIPALS)
+    assert response.passages == []
+    assert "couldn't find" in response.answer

engine/uv.lock

@@ -494,14 +494,14 @@ name = "cohere"
 version = "7.0.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "fastavro" },
-    { name = "httpx" },
-    { name = "pydantic" },
-    { name = "pydantic-core" },
-    { name = "requests" },
-    { name = "tokenizers" },
-    { name = "types-requests" },
-    { name = "typing-extensions" },
+    { name = "fastavro", marker = "sys_platform != 'emscripten'" },
+    { name = "httpx", marker = "sys_platform != 'emscripten'" },
+    { name = "pydantic", marker = "sys_platform != 'emscripten'" },
+    { name = "pydantic-core", marker = "sys_platform != 'emscripten'" },
+    { name = "requests", marker = "sys_platform != 'emscripten'" },
+    { name = "tokenizers", marker = "sys_platform != 'emscripten'" },
+    { name = "types-requests", marker = "sys_platform != 'emscripten'" },
+    { name = "typing-extensions", marker = "sys_platform != 'emscripten'" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/cf/3c/670631ee223d7b64d157dc3f309bf93bde65efe0bb1a8341d9b575f407d3/cohere-7.0.4.tar.gz", hash = "sha256:35b6a397d35ae6eafa1a02921f42c2a98309a990874533e5238efaf3426b6a21", size = 208794, upload-time = "2026-06-11T15:17:52.994Z" }
 wheels = [
@@ -855,6 +855,7 @@ dependencies = [
     { name = "pydantic-ai" },
     { name = "pydantic-ai-slim", extra = ["voyageai"] },
     { name = "pydantic-settings" },
+    { name = "python-docx" },
     { name = "python-dotenv" },
     { name = "sqlite-vec" },
     { name = "uvicorn" },
@@ -892,6 +893,7 @@ requires-dist = [
     { name = "pydantic-ai", specifier = ">=1.99.0,<2.0.0" },
     { name = "pydantic-ai-slim", extras = ["voyageai"], specifier = ">=1.99.0,<2.0.0" },
     { name = "pydantic-settings", specifier = ">=2.0.0" },
+    { name = "python-docx", specifier = ">=1.1.2" },
     { name = "python-dotenv", specifier = ">=1.2.1" },
     { name = "sqlite-vec", specifier = ">=0.1.6" },
     { name = "torch", marker = "sys_platform == 'linux' and extra == 'docparse'", specifier = ">=2.6.0", index = "https://download.pytorch.org/whl/cpu" },
@@ -4360,7 +4362,7 @@ name = "types-requests"
 version = "2.33.0.20260518"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "urllib3" },
+    { name = "urllib3", marker = "sys_platform != 'emscripten'" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/e0/01/c5a19253fe1ac159159ddf9a3a07cec8bb5e486ec4d9002ad2821da0e5d2/types_requests-2.33.0.20260518.tar.gz", hash = "sha256:df7bd3bfe0ca8402dfb841e7d9be714bb5578203283d66d7dc4ef69343449a5e", size = 24752, upload-time = "2026-05-18T06:07:37.966Z" }
 wheels = [

frontend/editor/public/locales/en-US/translation.toml

@@ -2986,6 +2986,38 @@ summary_one = "Ran 1 tool"
 summary_other = "Ran {{count}} tools"
 unknownTool = "Unknown tool"
 
+[chunkDocument]
+intro = "Turns a document into retrieval-ready chunks in three layers, so answers cite the right section instead of a random page."
+processorCallout = "To index automatically, add the 'Index into knowledge base' step to an ingestion policy in the"
+processorLink = "Processor"
+submit = "Prepare chunks"
+
+[chunkDocument.chunkSize]
+label = "Chunk size (characters)"
+
+[chunkDocument.error]
+failed = "Failed to chunk document"
+
+[chunkDocument.layers]
+chunk = "Structure-aware chunks: each carries its heading breadcrumb and page range"
+embed = "Ready to embed: exported as JSONL for any vector store"
+parse = "Layout-aware parse: headings, paragraphs, and tables are recognized as structure"
+
+[chunkDocument.mode]
+advanced = "Advanced"
+auto = "Auto"
+basic = "Basic"
+label = "Mode"
+
+[chunkDocument.overlap]
+label = "Overlap (characters)"
+
+[chunkDocument.results]
+title = "Chunks (JSONL)"
+
+[chunkDocument.settings]
+title = "Chunking settings"
+
 [cloudBadge]
 tooltip = "This operation will use your cloud credits"
 
@@ -3573,6 +3605,13 @@ tags = "automation,folder,scanning,watch folder,hot folder,automatic processing,
 [devSsoGuide]
 tags = "SSO,single sign-on,authentication,SAML,OAuth,OIDC,login,enterprise,identity provider,IdP"
 
+[docparse.intro]
+advancedOff = "Advanced parsing: off"
+advancedOn = "Advanced parsing: on"
+aiLayoutModel = "AI layout model"
+basicFallback = "Scanned documents fall back to basic text extraction - install the DocParse addon for layout AI"
+usesAi = "Uses AI"
+
 [dropdownList]
 searchPlaceholder = "Search..."
 
@@ -3700,6 +3739,56 @@ _value = "Error"
 dismissAllErrors = "Dismiss All Errors"
 generic = "An error occurred"
 
+[extractFields]
+intro = "Describe the fields you need and AI reads the document and returns each value with a confidence score and a citation you can verify."
+submit = "Extract fields"
+
+[extractFields.error]
+failed = "Failed to extract fields"
+
+[extractFields.fields]
+add = "Add field"
+description = "Description"
+descriptionPlaceholder = "What to look for"
+label = "Fields to extract"
+name = "Name"
+namePlaceholder = "invoice_number"
+remove = "Remove field"
+type = "Type"
+
+[extractFields.instructions]
+label = "Instructions"
+placeholder = "e.g. Amounts are in EUR unless stated otherwise"
+
+[extractFields.mode]
+advanced = "Advanced"
+auto = "Auto"
+basic = "Basic"
+label = "Mode"
+
+[extractFields.presets]
+contract = "Contract"
+custom = "Custom"
+invoice = "Invoice"
+label = "Preset template"
+purchaseOrder = "Purchase order"
+receipt = "Receipt"
+
+[extractFields.results]
+title = "Extraction report"
+
+[extractFields.resultsPanel]
+title = "Extracted fields"
+
+[extractFields.settings]
+title = "Extraction schema"
+
+[extractFields.suggest]
+button = "Suggest fields (AI)"
+failed = "Could not suggest fields"
+failedBody = "The document could not be analyzed. Add fields manually or try again."
+needsFile = "Select a file first to suggest fields"
+
 [extractImages]
 allowDuplicates = "Save duplicate images"
 selectText = "Select image format to convert extracted images to"
@@ -4183,6 +4272,24 @@ upload = "Upload"
 uploadFile = "Upload File"
 uploadFiles = "Upload Files"
 
+[fillTemplate]
+hint = "The input file must be a .docx template (not a PDF). Each placeholder in the template is replaced with the matching JSON value."
+intro = "Replaces the placeholders in a Word (.docx) template with your JSON data and returns the filled document - deterministic, no AI involved."
+submit = "Fill template"
+
+[fillTemplate.data]
+invalid = "Enter a valid JSON object"
+label = "Data (JSON)"
+
+[fillTemplate.error]
+failed = "Failed to fill template"
+
+[fillTemplate.results]
+title = "Filled document"
+
+[fillTemplate.settings]
+title = "Template data"
+
 [firstLogin]
 allFieldsRequired = "All fields are required"
 changePassword = "Change Password"
@@ -4522,6 +4629,11 @@ desc = "Change document restrictions and permissions"
 tags = "permissions,restrictions,rights,access control,allow,deny,printing,copying,editing,modify permissions,security settings,user rights"
 title = "Change Permissions"
 
+[home.chunkDocument]
+desc = "Layout-aware parse to structure-aware chunks with heading breadcrumbs and page ranges, ready to embed"
+tags = "chunk,RAG,prepare,split text,segments,embedding,vector,ingest,LLM,retrieval,JSONL,overlap,knowledge base,index"
+title = "Prepare for RAG"
+
 [home.compare]
 desc = "Compares and shows the differences between 2 PDF Documents"
 tags = "difference,compare,diff,compare PDFs,compare documents,find differences,show differences,changes,what changed,track changes,revisions,version compare,side by side,contrast,delta"
@@ -4567,6 +4679,11 @@ desc = "Add or edit bookmarks and table of contents in PDF documents"
 tags = "bookmarks,contents,edit,table of contents,TOC,outline,navigation,chapters,sections,add bookmarks,edit bookmarks,PDF outline"
 title = "Edit Table of Contents"
 
+[home.extractFields]
+desc = "Pull typed fields out of a document with confidence scores and citations"
+tags = "extract,fields,schema,structured data,invoice,form data,key value,confidence,citations,capture,parse"
+title = "Extract Fields"
+
 [home.extractImages]
 desc = "Extracts all images from a PDF and saves them to zip"
 tags = "pull,save,export,extract images,get images,save images,export images,extract photos,extract pictures,pull images,download images,rip images,extract graphics,save photos"
@@ -4577,6 +4694,11 @@ desc = "Extract specific pages from a PDF document"
 tags = "pull,select,copy,extract,extract pages,get pages,pull out,save pages,export pages,copy pages,select pages,specific pages"
 title = "Extract Pages"
 
+[home.fillTemplate]
+desc = "Fill a DOCX template's placeholders from JSON data"
+tags = "template,DOCX,fill,merge fields,mail merge,generate document,placeholders,letters,contracts,Word"
+title = "Fill Template"
+
 [home.flatten]
 desc = "Remove all interactive elements and forms from a PDF"
 tags = "simplify,remove,interactive,flatten,flatten form,remove form fields,make static,finalize form,lock form,disable editing,convert to image,non-editable"
@@ -4630,6 +4752,11 @@ desc = "Merge multiple pages of a PDF document into a single page"
 tags = "layout,arrange,combine,N-up,2-up,4-up,multiple per page,pages per sheet,layout pages,tile,grid layout,multi-page layout,combine on page,handout"
 title = "Multi-Page Layout"
 
+[home.parseDocument]
+desc = "Layout-aware parsing to structured JSON or Markdown, with optional OCR"
+tags = "parse,layout,structure,blocks,markdown,JSON,docling,OCR,scan,document understanding,convert"
+title = "Parse Document"
+
 [home.pdfCommentAgent]
 desc = "Ask AI to annotate a PDF with sticky-note comments based on your prompt"
 tags = "AI,agent,comment,annotate,sticky note,review,feedback,notes"
@@ -4743,6 +4870,11 @@ desc = "Adds signature to PDF by drawing, text or image"
 tags = "signature,autograph,e-sign,electronic signature,digital signature,sign document,approval,signoff,authorize,endorse,ink signature,handwriting"
 title = "Sign"
 
+[home.smartSplit]
+desc = "Split a PDF into sub-documents using a natural-language boundary rule"
+tags = "split,smart,boundaries,separate,invoices,batches,content split,divide,rules,auto split"
+title = "Smart Split"
+
 [home.split]
 desc = "Split PDFs into multiple documents"
 tags = "divide,separate,break,split,extract pages,separate pages,divide document,break apart,separate files,unbind,split by page,divide by chapter"
@@ -5501,6 +5633,33 @@ title = "Page Ranges"
 bullet1 = "<strong>all</strong> → selects all pages"
 title = "Special Keywords"
 
+[parseDocument]
+intro = "Reads the document's layout - headings, paragraphs, tables - and turns it into clean structured JSON or Markdown you can feed to other systems."
+submit = "Parse document"
+
+[parseDocument.error]
+failed = "Failed to parse document"
+
+[parseDocument.mode]
+advanced = "Advanced"
+auto = "Auto"
+basic = "Basic"
+label = "Mode"
+
+[parseDocument.outputFormat]
+json = "JSON"
+label = "Output format"
+markdown = "Markdown"
+
+[parseDocument.results]
+title = "Parsed output"
+
+[parseDocument.settings]
+title = "Parse settings"
+
+[parseDocument.withOcr]
+label = "Apply OCR to scanned pages (recommended)"
+
 [payg.activity]
 docs = "docs"
 empty = "No billable activity yet this period."
@@ -7877,6 +8036,10 @@ label = "Classification"
 desc = "Enforce HIPAA, GDPR, SOC 2, or FedRAMP requirements on every document."
 label = "Compliance"
 
+[portal.policies.categories.docIntelligence]
+desc = "Extract the structured fields you describe from every document, with confidence scores and citations."
+label = "Document intelligence"
+
 [portal.policies.categories.ingestion]
 desc = "Normalize incoming documents - OCR scans, flatten forms - and index them into the searchable knowledge base, or export them as a clean corpus for your own systems."
 label = "Ingestion"
@@ -7917,6 +8080,18 @@ onViolation = "When non-compliant"
 1 = "Enforce action"
 2 = "Audit trail"
 
+[portal.policies.config.docIntelligence]
+summary = "Extracts the fields you describe from every document, with confidence and citations."
+
+[portal.policies.config.docIntelligence.rules]
+0 = "Extract the fields you describe, with confidence and citations"
+
+[portal.policies.config.extractFields.fields]
+instructions = "Extraction guidance (optional)"
+instructionsHelp = "Plain-language hints for tricky fields, e.g. \"the invoice number is top right\"."
+schema = "Fields to extract (JSON Schema)"
+schemaHelp = "A JSON Schema object describing the fields. Each extracted value carries a confidence score and a citation."
+
 [portal.policies.config.ingestion]
 summary = "Normalizes incoming documents and indexes them into the searchable knowledge base, with optional markdown/JSONL export."
 
@@ -8036,6 +8211,7 @@ addWatermark = "Watermark"
 autoRedact = "Redact PII"
 classifyAndLabel = "Classify"
 compressPdf = "Compress"
+extractFields = "Extract fields"
 flatten = "Flatten"
 ocrPdf = "OCR"
 ragIngest = "Ingest into knowledge base"
@@ -8270,6 +8446,10 @@ label = "Classify the document"
 desc = "Compresses the document to a smaller file size."
 label = "Reduce file size"
 
+[portal.policies.wizard.capability.extractFields]
+desc = "Pulls the values you describe - like invoice numbers or dates - out of every document, each with a confidence score and a citation back to the page."
+label = "Extract structured fields"
+
 [portal.policies.wizard.capability.flatten]
 desc = "Merges form fields and annotations into the page so they can't be edited."
 label = "Flatten the document"
@@ -10284,6 +10464,26 @@ medium = "Medium"
 small = "Small"
 x-large = "X-Large"
 
+[smartSplit]
+intro = "Describe where sub-documents start in plain language and AI reads the content to find those boundaries - no page numbers needed."
+submit = "Split document"
+
+[smartSplit.error]
+failed = "Failed to split document"
+
+[smartSplit.maxParts]
+label = "Maximum parts"
+
+[smartSplit.results]
+title = "Split documents"
+
+[smartSplit.rule]
+label = "Split rule"
+placeholder = "e.g. Start a new document at every invoice header"
+
+[smartSplit.settings]
+title = "Split settings"
+
 [split]
 resultsTitle = "Split Results"
 selectMethod = "Select a split method"
@@ -10750,6 +10950,7 @@ standardTools = "Standard Tools"
 advancedFormatting = "Advanced Formatting"
 automation = "Automation"
 developerTools = "Developer Tools"
+documentIntelligence = "Document intelligence"
 documentReview = "Document Review"
 documentSecurity = "Document Security"
 extraction = "Extraction"

frontend/editor/src/core/data/toolsTaxonomy.ts

@@ -22,6 +22,7 @@ import ViewAgendaRoundedIcon from "@mui/icons-material/ViewAgendaRounded";
 import FileDownloadRoundedIcon from "@mui/icons-material/FileDownloadRounded";
 import DeleteSweepRoundedIcon from "@mui/icons-material/DeleteSweepRounded";
 import SmartToyRoundedIcon from "@mui/icons-material/SmartToyRounded";
+import AutoAwesomeRoundedIcon from "@mui/icons-material/AutoAwesomeRounded";
 import BuildRoundedIcon from "@mui/icons-material/BuildRounded";
 import TuneRoundedIcon from "@mui/icons-material/TuneRounded";
 import CodeRoundedIcon from "@mui/icons-material/CodeRounded";
@@ -34,6 +35,7 @@ export enum SubcategoryId {
   VERIFICATION = "verification",
   DOCUMENT_REVIEW = "documentReview",
   PAGE_FORMATTING = "pageFormatting",
+  DOCUMENT_INTELLIGENCE = "documentIntelligence",
   EXTRACTION = "extraction",
   REMOVAL = "removal",
   AUTOMATION = "automation",
@@ -95,6 +97,7 @@ export const SUBCATEGORY_ORDER: SubcategoryId[] = [
   SubcategoryId.VERIFICATION,
   SubcategoryId.DOCUMENT_REVIEW,
   SubcategoryId.PAGE_FORMATTING,
+  SubcategoryId.DOCUMENT_INTELLIGENCE,
   SubcategoryId.EXTRACTION,
   SubcategoryId.REMOVAL,
   SubcategoryId.AUTOMATION,
@@ -109,6 +112,7 @@ export const SUBCATEGORY_COLOR_MAP: Record<SubcategoryId, string> = {
   [SubcategoryId.VERIFICATION]: "var(--category-color-verification)", // Orange
   [SubcategoryId.DOCUMENT_REVIEW]: "var(--category-color-general)", // Blue
   [SubcategoryId.PAGE_FORMATTING]: "var(--category-color-formatting)", // Purple
+  [SubcategoryId.DOCUMENT_INTELLIGENCE]: "var(--category-color-automation)", // Pink
   [SubcategoryId.EXTRACTION]: "var(--category-color-extraction)", // Cyan
   [SubcategoryId.REMOVAL]: "var(--category-color-removal)", // Red
   [SubcategoryId.AUTOMATION]: "var(--category-color-automation)", // Pink
@@ -131,6 +135,8 @@ export const getSubcategoryIcon = (
       return React.createElement(RateReviewRoundedIcon);
     case SubcategoryId.PAGE_FORMATTING:
       return React.createElement(ViewAgendaRoundedIcon);
+    case SubcategoryId.DOCUMENT_INTELLIGENCE:
+      return React.createElement(AutoAwesomeRoundedIcon);
     case SubcategoryId.EXTRACTION:
       return React.createElement(FileDownloadRoundedIcon);
     case SubcategoryId.REMOVAL:

frontend/editor/src/core/hooks/useDocparseEnabled.ts

@@ -0,0 +1,10 @@
+import { useAppConfig } from "@app/contexts/AppConfigContext";
+
+/**
+ * Whether the DocParse layer is enabled, per the backend's app-config.
+ * Gates the DocParse tools' visibility; flavors may shadow this hook.
+ */
+export function useDocparseEnabled(): boolean {
+  const { config } = useAppConfig();
+  return Boolean(config?.docparseEnabled);
+}

frontend/editor/src/core/tests/stubbed/docparse-walkthrough.spec.ts

@@ -0,0 +1,165 @@
+import { test, expect } from "@app/tests/helpers/stub-test-base";
+import type { Page, Route } from "@playwright/test";
+import path from "node:path";
+
+/** DocParse walkthrough: the five workbench tools.
+ * Dumps PNGs to screenshots/docparse; light + dark per view, RTL spot checks. */
+
+const SCREENSHOTS_DIR = path.resolve(process.cwd(), "screenshots", "docparse");
+
+function shotPath(name: string): string {
+  return path.join(SCREENSHOTS_DIR, `${name}.png`);
+}
+
+async function settle(page: Page, ms = 400): Promise<void> {
+  await page.waitForTimeout(ms);
+}
+
+async function stubApis(page: Page): Promise<void> {
+  // Narrow fallbacks only: a blanket /api/v1/** would out-rank the stub
+  // fixture's own /auth/me route (last-registered wins) and break the session.
+  await page.route("**/api/v1/policies/**", (route: Route) =>
+    route.fulfill({ json: [] }),
+  );
+  await page.route("**/api/v1/proprietary/ui-data/**", (route: Route) =>
+    route.fulfill({ json: [] }),
+  );
+  // enableLogin true: the portal only builds a session when login mode is on
+  // (matches live behavior; with login off the portal shows its login screen).
+  const configPayload = {
+    appVersion: "test",
+    enableLogin: true,
+    isAdmin: true,
+    languages: ["en-US"],
+    defaultLocale: "en-US",
+    aiEngineEnabled: true,
+    docparseEnabled: true,
+    docparseAdvanced: true,
+    storageEnabled: false,
+    premiumEnabled: true,
+    runningProOrHigher: true,
+  };
+  await page.route("**/api/v1/config/app-config", (route: Route) =>
+    route.fulfill({ json: configPayload }),
+  );
+  // The auth layer decides login mode from public-config; keep it in sync.
+  await page.route("**/api/v1/config/public-config", (route: Route) =>
+    route.fulfill({
+      json: { enableLogin: true, languages: ["en-US"], defaultLocale: "en-US" },
+    }),
+  );
+  await page.route(
+    "**/api/v1/config/endpoints-availability**",
+    (route: Route) => route.fulfill({ json: {} }),
+  );
+  await page.route("**/api/v1/config/endpoint-enabled**", (route: Route) =>
+    route.fulfill({ json: { enabled: true } }),
+  );
+  // DocparseToolIntro probes live capabilities for its tier badges.
+  await page.route("**/api/v1/docparse/capabilities", (route: Route) =>
+    route.fulfill({
+      json: {
+        enabled: true,
+        mode: "auto",
+        advancedInstalled: true,
+        engineReachable: true,
+        doclingVersion: "2.116.0",
+      },
+    }),
+  );
+}
+
+async function enableDarkMode(page: Page): Promise<void> {
+  await page.addInitScript(() => {
+    localStorage.setItem("mantine-color-scheme", "dark");
+    localStorage.setItem("mantine-color-scheme-value", "dark");
+  });
+  await page.emulateMedia({ colorScheme: "dark" });
+}
+
+async function enableRtl(page: Page): Promise<void> {
+  await page.addInitScript(() => {
+    localStorage.setItem("i18nextLng", "ar-AR");
+    localStorage.setItem("stirling-language", "ar-AR");
+    localStorage.setItem("stirling-language-source", "user");
+    const applyDir = () => {
+      document.documentElement.setAttribute("dir", "rtl");
+      document.documentElement.setAttribute("lang", "ar-AR");
+    };
+    if (document.documentElement) applyDir();
+    else document.addEventListener("DOMContentLoaded", applyDir);
+  });
+}
+
+const TOOLS = [
+  { id: "parseDocument", url: "/parse-document", waitText: /Parse/i },
+  { id: "extractFields", url: "/extract-fields", waitText: /Extract/i },
+  { id: "smartSplit", url: "/smart-split", waitText: /Split/i },
+  { id: "chunkDocument", url: "/chunk-document", waitText: /Chunk/i },
+  { id: "fillTemplate", url: "/fill-template", waitText: /Template|Fill/i },
+];
+
+async function openTool(page: Page, url: string): Promise<void> {
+  await page.goto(url, { waitUntil: "domcontentloaded" });
+  await expect(page.locator("body").first()).not.toBeEmpty();
+  // The tool panel is the left rail; give lazy chunks a moment.
+  await settle(page, 900);
+}
+
+test.describe("DocParse walkthrough", () => {
+  test.use({
+    autoGoto: false,
+    viewport: { width: 1600, height: 900 },
+    seedJwt: true,
+  });
+
+  // ─── Editor tools, light ──────────────────────────────────────────────────
+  for (const [i, tool] of TOOLS.entries()) {
+    test(`t${i}_${tool.id}_light`, async ({ page }) => {
+      await stubApis(page);
+      await openTool(page, tool.url);
+      await page.screenshot({ path: shotPath(`0${i + 1}_${tool.id}_light`) });
+    });
+
+    test(`t${i}_${tool.id}_dark`, async ({ page }) => {
+      await enableDarkMode(page);
+      await stubApis(page);
+      await openTool(page, tool.url);
+      await page.screenshot({ path: shotPath(`0${i + 1}_${tool.id}_dark`) });
+    });
+  }
+
+  // ─── Extract Fields with builder rows filled ─────────────────────────────
+  for (const theme of ["light", "dark"] as const) {
+    test(`extract_fields_populated_${theme}`, async ({ page }) => {
+      if (theme === "dark") await enableDarkMode(page);
+      await stubApis(page);
+      await openTool(page, "/extract-fields");
+      // Fill the first schema-builder row when present; tolerate layout drift.
+      const nameInput = page.getByPlaceholder(/name/i).first();
+      if (await nameInput.isVisible().catch(() => false)) {
+        await nameInput.fill("invoice_number");
+        const addButton = page.getByRole("button", { name: /add/i }).first();
+        if (await addButton.isVisible().catch(() => false)) {
+          await addButton.click();
+          const second = page.getByPlaceholder(/name/i).nth(1);
+          if (await second.isVisible().catch(() => false)) {
+            await second.fill("total_due");
+          }
+        }
+      }
+      await settle(page);
+      await page.screenshot({
+        path: shotPath(`06_extract_fields_populated_${theme}`),
+      });
+    });
+  }
+
+  // ─── RTL spot checks ──────────────────────────────────────────────────────
+  test("rtl_extract_fields", async ({ page }) => {
+    await enableRtl(page);
+    await stubApis(page);
+    await openTool(page, "/extract-fields");
+    await page.screenshot({ path: shotPath("11_extract_fields_rtl") });
+  });
+});

frontend/editor/src/core/types/appConfig.ts

@@ -62,6 +62,7 @@ export interface AppConfig {
   timestampCustomTsaUrls?: string[];
   timestampTsaPresets?: { label: string; url: string }[];
   aiEngineEnabled?: boolean;
+  docparseEnabled?: boolean;
 }
 
 export type AppConfigBootstrapMode = "blocking" | "non-blocking";

frontend/editor/src/portal/api/policies.ts

@@ -141,7 +141,8 @@ export const ENDPOINT_LABELS: Partial<
   Record<
     | ToolEndpoint
     | "/api/v1/ai/tools/classify-and-label"
-    | "/api/v1/docparse/rag-ingest",
+    | "/api/v1/docparse/rag-ingest"
+    | "/api/v1/docparse/extract-fields",
     string
   >
 > = {
@@ -154,6 +155,7 @@ export const ENDPOINT_LABELS: Partial<
   "/api/v1/ai/tools/classify-and-label":
     "portal.policies.endpoints.classifyAndLabel",
   "/api/v1/docparse/rag-ingest": "portal.policies.endpoints.ragIngest",
+  "/api/v1/docparse/extract-fields": "portal.policies.endpoints.extractFields",
 };
 
 export function humanizeEndpoint(
@@ -180,6 +182,13 @@ const DEFAULT_PII_PATTERNS: string[] = [
 
 /** `label`/`desc` values are i18n keys — render with t(). */
 export const POLICY_CATEGORIES: PolicyCategory[] = [
+  // First on purpose: document intelligence is the processor's flagship flow.
+  {
+    id: "docIntelligence",
+    label: "portal.policies.categories.docIntelligence.label",
+    tone: "blue",
+    desc: "portal.policies.categories.docIntelligence.desc",
+  },
   {
     id: "ingestion",
     label: "portal.policies.categories.ingestion.label",
@@ -228,6 +237,13 @@ export const POLICY_CATEGORIES: PolicyCategory[] = [
  * stay as stable values (translating them would corrupt saved configs).
  */
 export const POLICY_CONFIG: Record<string, PolicyConfigDef> = {
+  docIntelligence: {
+    summary: "portal.policies.config.docIntelligence.summary",
+    rules: ["portal.policies.config.docIntelligence.rules.0"],
+    scopeLabel: "portal.policies.config.scopeAll",
+    defaultOperations: [policyStep("extractFields")],
+    fields: [],
+  },
   ingestion: {
     summary: "portal.policies.config.ingestion.summary",
     rules: [

frontend/editor/src/portal/components/policies/PolicyExtractFieldsConfig.tsx

@@ -0,0 +1,54 @@
+import { useTranslation } from "react-i18next";
+import { FormField, Input } from "@app/ui";
+
+/** Configures the extract-fields step: the schema and optional guidance. */
+export interface ExtractFieldsStepParams {
+  fieldsSchema: string;
+  mode: string;
+  instructions: string;
+}
+
+interface PolicyExtractFieldsConfigProps {
+  parameters: ExtractFieldsStepParams;
+  onChange: (parameters: ExtractFieldsStepParams) => void;
+}
+
+export function PolicyExtractFieldsConfig({
+  parameters,
+  onChange,
+}: PolicyExtractFieldsConfigProps) {
+  const { t } = useTranslation();
+
+  return (
+    <div className="portal-policies__capability-config">
+      <FormField
+        label={t("portal.policies.config.extractFields.fields.schema")}
+        helperText={t("portal.policies.config.extractFields.fields.schemaHelp")}
+      >
+        <textarea
+          className="portal-sources__connection-textarea"
+          rows={4}
+          value={parameters.fieldsSchema ?? ""}
+          placeholder='{"type": "object", "properties": {"invoice_number": {"type": "string"}}}'
+          onChange={(e) =>
+            onChange({ ...parameters, fieldsSchema: e.target.value })
+          }
+        />
+      </FormField>
+      <FormField
+        label={t("portal.policies.config.extractFields.fields.instructions")}
+        helperText={t(
+          "portal.policies.config.extractFields.fields.instructionsHelp",
+        )}
+      >
+        <Input
+          inputSize="sm"
+          value={parameters.instructions ?? ""}
+          onChange={(e) =>
+            onChange({ ...parameters, instructions: e.target.value })
+          }
+        />
+      </FormField>
+    </div>
+  );
+}

frontend/editor/src/portal/components/policies/PolicySetupWizard.tsx

@@ -40,6 +40,7 @@ import { PolicyCategoryBadge } from "@portal/components/policies/PolicyCategoryI
 import { PolicyRedactConfig } from "@app/components/policies/PolicyRedactConfig";
 import { PolicyWatermarkConfig } from "@app/components/policies/PolicyWatermarkConfig";
 import { PolicyPurviewConfig } from "@portal/components/policies/PolicyPurviewConfig";
+import { PolicyExtractFieldsConfig } from "@portal/components/policies/PolicyExtractFieldsConfig";
 import { PolicyRagIngestConfig } from "@portal/components/policies/PolicyRagIngestConfig";
 import { ClassificationLabelsSection } from "@portal/components/policies/ClassificationLabelsSection";
 import "@portal/views/Policies.css";
@@ -98,6 +99,8 @@ const DISABLED_BY_DEFAULT = new Set<PolicyToolId>([
   "purviewApplyLabel",
   "purviewReadLabel",
   "externalApiCall",
+  // Needs a fields schema before a run can succeed.
+  "extractFields",
 ]);
 
 // Steps that cannot work without a Purview tenant connection, so they are hidden entirely until one
@@ -193,6 +196,13 @@ const CAPABILITY_META: Record<
     descEn:
       "Hands the document to a system you have connected, and records what it answered.",
   },
+  extractFields: {
+    labelKey: "portal.policies.wizard.capability.extractFields.label",
+    labelEn: "Extract structured fields",
+    descKey: "portal.policies.wizard.capability.extractFields.desc",
+    descEn:
+      "Pulls the values you describe - like invoice numbers or dates - out of every document, each with a confidence score and a citation back to the page.",
+  },
   ragIngest: {
     labelKey: "portal.policies.wizard.capability.ragIngest.label",
     labelEn: "Index into the knowledge base",
@@ -562,6 +572,14 @@ function PolicySetupWizardBody({
                             }
                           />
                         )}
+                        {tl.toolId === "extractFields" && (
+                          <PolicyExtractFieldsConfig
+                            parameters={tl.params}
+                            onChange={(params) =>
+                              setToolParams("extractFields", params)
+                            }
+                          />
+                        )}
                         {tl.toolId === "ragIngest" && (
                           <PolicyRagIngestConfig
                             parameters={tl.params}

frontend/editor/src/proprietary/components/tools/docparse/ChunkDocumentSettings.tsx

@@ -0,0 +1,91 @@
+import { useTranslation } from "react-i18next";
+import { Anchor, List, NumberInput, Select, Stack, Text } from "@mantine/core";
+import type { ToolAutomationSettingsProps } from "@app/hooks/tools/shared/toolOperationTypes";
+import type { ChunkDocumentParameters } from "@app/hooks/tools/chunkDocument/useChunkDocumentParameters";
+import type { DocparseMode } from "@app/hooks/tools/parseDocument/useParseDocumentParameters";
+import DocparseToolIntro from "@app/components/tools/docparse/DocparseToolIntro";
+
+const ChunkDocumentSettings = ({
+  parameters,
+  onParameterChange,
+  disabled,
+}: ToolAutomationSettingsProps<ChunkDocumentParameters>) => {
+  const { t } = useTranslation();
+
+  return (
+    <Stack gap="sm">
+      <DocparseToolIntro
+        description={t(
+          "chunkDocument.intro",
+          "Turns a document into retrieval-ready chunks in three layers, so answers cite the right section instead of a random page.",
+        )}
+        aiBadge="layout"
+      />
+      <List type="ordered" size="sm" spacing={4}>
+        <List.Item>
+          {t(
+            "chunkDocument.layers.parse",
+            "Layout-aware parse: headings, paragraphs, and tables are recognized as structure",
+          )}
+        </List.Item>
+        <List.Item>
+          {t(
+            "chunkDocument.layers.chunk",
+            "Structure-aware chunks: each carries its heading breadcrumb and page range",
+          )}
+        </List.Item>
+        <List.Item>
+          {t(
+            "chunkDocument.layers.embed",
+            "Ready to embed: exported as JSONL for any vector store",
+          )}
+        </List.Item>
+      </List>
+      <NumberInput
+        label={t("chunkDocument.chunkSize.label", "Chunk size (characters)")}
+        value={parameters.chunkSize}
+        onChange={(value) =>
+          onParameterChange("chunkSize", typeof value === "number" ? value : 0)
+        }
+        min={1}
+        disabled={disabled}
+      />
+      <NumberInput
+        label={t("chunkDocument.overlap.label", "Overlap (characters)")}
+        value={parameters.overlap}
+        onChange={(value) =>
+          onParameterChange("overlap", typeof value === "number" ? value : 0)
+        }
+        min={0}
+        disabled={disabled}
+      />
+      <Select
+        label={t("chunkDocument.mode.label", "Mode")}
+        value={parameters.mode}
+        onChange={(value) =>
+          onParameterChange("mode", (value ?? "auto") as DocparseMode)
+        }
+        data={[
+          { value: "auto", label: t("chunkDocument.mode.auto", "Auto") },
+          { value: "basic", label: t("chunkDocument.mode.basic", "Basic") },
+          {
+            value: "advanced",
+            label: t("chunkDocument.mode.advanced", "Advanced"),
+          },
+        ]}
+        disabled={disabled}
+      />
+      <Text size="xs" c="dimmed">
+        {t(
+          "chunkDocument.processorCallout",
+          "To index automatically, add the 'Index into knowledge base' step to an ingestion policy in the",
+        )}{" "}
+        <Anchor href="/processor/policies" size="xs">
+          {t("chunkDocument.processorLink", "Processor")}
+        </Anchor>
+      </Text>
+    </Stack>
+  );
+};
+
+export default ChunkDocumentSettings;

frontend/editor/src/proprietary/components/tools/docparse/DocparseToolIntro.module.css

@@ -0,0 +1,30 @@
+.card {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+  padding: 0.625rem 0.75rem;
+  border: 1px solid var(--c-border-subtle);
+  border-radius: 8px;
+  background: var(--c-surface-sunken);
+}
+
+.description {
+  margin: 0;
+  font-size: 0.8125rem;
+  line-height: 1.45;
+  color: var(--c-text-subtle);
+}
+
+.badges {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 0.375rem;
+}
+
+.fallback {
+  margin: 0;
+  font-size: 0.75rem;
+  line-height: 1.4;
+  color: var(--c-text-subtle);
+  font-style: italic;
+}

frontend/editor/src/proprietary/components/tools/docparse/DocparseToolIntro.tsx

@@ -0,0 +1,62 @@
+import { useTranslation } from "react-i18next";
+import { Badge } from "@mantine/core";
+import { useDocparseCapabilities } from "@app/hooks/useDocparseCapabilities";
+import styles from "@app/components/tools/docparse/DocparseToolIntro.module.css";
+
+interface DocparseToolIntroProps {
+  /** 1-2 sentence "how this works" copy for the tool, already translated. */
+  description: string;
+  /**
+   * Which AI badge fits the tool: "llm" tools always call the language model;
+   * "layout" tools use the layout AI model only when advanced parsing is on.
+   */
+  aiBadge?: "llm" | "layout";
+  /** Hide the scanned-docs fallback note where it cannot apply (DOCX input). */
+  showFallbackNote?: boolean;
+}
+
+/** Compact "how this works" card at the top of every DocParse settings panel. */
+const DocparseToolIntro = ({
+  description,
+  aiBadge,
+  showFallbackNote = true,
+}: DocparseToolIntroProps) => {
+  const { t } = useTranslation();
+  const { capabilities } = useDocparseCapabilities();
+  const advanced = capabilities ? capabilities.advancedInstalled : null;
+
+  return (
+    <div className={styles.card}>
+      <p className={styles.description}>{description}</p>
+      <div className={styles.badges}>
+        {aiBadge === "llm" && (
+          <Badge size="sm" variant="light" color="grape">
+            {t("docparse.intro.usesAi", "Uses AI")}
+          </Badge>
+        )}
+        {aiBadge === "layout" && advanced === true && (
+          <Badge size="sm" variant="light" color="grape">
+            {t("docparse.intro.aiLayoutModel", "AI layout model")}
+          </Badge>
+        )}
+        {advanced !== null && (
+          <Badge size="sm" variant="light" color={advanced ? "teal" : "gray"}>
+            {advanced
+              ? t("docparse.intro.advancedOn", "Advanced parsing: on")
+              : t("docparse.intro.advancedOff", "Advanced parsing: off")}
+          </Badge>
+        )}
+      </div>
+      {advanced === false && showFallbackNote && (
+        <p className={styles.fallback}>
+          {t(
+            "docparse.intro.basicFallback",
+            "Scanned documents fall back to basic text extraction - install the DocParse addon for layout AI",
+          )}
+        </p>
+      )}
+    </div>
+  );
+};
+
+export default DocparseToolIntro;

frontend/editor/src/proprietary/components/tools/docparse/ExtractFieldsSettings.module.css

@@ -0,0 +1,36 @@
+.schema {
+  display: flex;
+  flex-direction: column;
+  gap: 0.5rem;
+}
+
+/* Two-line card per field: the side panel is too narrow for three columns. */
+.row {
+  display: grid;
+  grid-template-columns: minmax(0, 1fr) 5.75rem auto;
+  grid-template-areas:
+    "name type remove"
+    "desc desc desc";
+  column-gap: 0.5rem;
+  row-gap: 0.375rem;
+  align-items: center;
+  padding: 0.5rem;
+  border: 1px solid var(--c-border-subtle);
+  border-radius: 0.5rem;
+}
+
+.nameInput {
+  grid-area: name;
+}
+
+.typeSelect {
+  grid-area: type;
+}
+
+.removeButton {
+  grid-area: remove;
+}
+
+.descriptionInput {
+  grid-area: desc;
+}

frontend/editor/src/proprietary/components/tools/docparse/ExtractFieldsSettings.tsx

@@ -0,0 +1,238 @@
+import { useState } from "react";
+import { useTranslation } from "react-i18next";
+import { Group, Select, Stack, Text, TextInput, Textarea } from "@mantine/core";
+import { Button } from "@app/ui";
+import DeleteOutlineRoundedIcon from "@mui/icons-material/DeleteOutlineRounded";
+import AddRoundedIcon from "@mui/icons-material/AddRounded";
+import AutoAwesomeRoundedIcon from "@mui/icons-material/AutoAwesomeRounded";
+import { alert } from "@app/components/toast";
+import type { ToolAutomationSettingsProps } from "@app/hooks/tools/shared/toolOperationTypes";
+import type { ExtractFieldsParameters } from "@app/hooks/tools/extractFields/useExtractFieldsParameters";
+import type { DocparseMode } from "@app/hooks/tools/shared/docparseTypes";
+import {
+  FIELD_TYPES,
+  emptyFieldRow,
+  type FieldRow,
+  type FieldType,
+} from "@app/hooks/tools/extractFields/fieldsSchema";
+import {
+  CUSTOM_PRESET,
+  FIELD_PRESET_IDS,
+  matchPreset,
+  presetRows,
+  type FieldPresetId,
+} from "@app/hooks/tools/extractFields/fieldsSchemaPresets";
+import { requestSuggestedFields } from "@app/hooks/tools/extractFields/suggestSchema";
+import DocparseToolIntro from "@app/components/tools/docparse/DocparseToolIntro";
+import styles from "@app/components/tools/docparse/ExtractFieldsSettings.module.css";
+
+interface ExtractFieldsSettingsProps extends ToolAutomationSettingsProps<ExtractFieldsParameters> {
+  /** The selected input file; enables the AI schema suggestion. */
+  selectedFile?: File | null;
+}
+
+/** Schema builder: rows of name/type/description plus free-form instructions. */
+const ExtractFieldsSettings = ({
+  parameters,
+  onParameterChange,
+  disabled,
+  selectedFile,
+}: ExtractFieldsSettingsProps) => {
+  const { t } = useTranslation();
+  const [suggesting, setSuggesting] = useState(false);
+
+  const setRow = (index: number, patch: Partial<FieldRow>) => {
+    const fields = parameters.fields.map((row, i) =>
+      i === index ? { ...row, ...patch } : row,
+    );
+    onParameterChange("fields", fields);
+  };
+
+  const removeRow = (index: number) => {
+    const fields = parameters.fields.filter((_, i) => i !== index);
+    onParameterChange("fields", fields.length > 0 ? fields : [emptyFieldRow()]);
+  };
+
+  const applyPreset = (value: string | null) => {
+    if (value === CUSTOM_PRESET) {
+      onParameterChange("fields", [emptyFieldRow()]);
+    } else if (value) {
+      onParameterChange("fields", presetRows(value as FieldPresetId));
+    }
+  };
+
+  const suggestFields = async () => {
+    if (!selectedFile || suggesting) return;
+    setSuggesting(true);
+    try {
+      const rows = await requestSuggestedFields(selectedFile);
+      onParameterChange("fields", rows.length > 0 ? rows : [emptyFieldRow()]);
+    } catch {
+      alert({
+        alertType: "error",
+        title: t("extractFields.suggest.failed", "Could not suggest fields"),
+        body: t(
+          "extractFields.suggest.failedBody",
+          "The document could not be analyzed. Add fields manually or try again.",
+        ),
+        expandable: false,
+      });
+    } finally {
+      setSuggesting(false);
+    }
+  };
+
+  const presetLabels: Record<FieldPresetId, string> = {
+    invoice: t("extractFields.presets.invoice", "Invoice"),
+    receipt: t("extractFields.presets.receipt", "Receipt"),
+    contract: t("extractFields.presets.contract", "Contract"),
+    purchaseOrder: t("extractFields.presets.purchaseOrder", "Purchase order"),
+  };
+
+  return (
+    <Stack gap="sm">
+      <DocparseToolIntro
+        description={t(
+          "extractFields.intro",
+          "Describe the fields you need and AI reads the document and returns each value with a confidence score and a citation you can verify.",
+        )}
+        aiBadge="llm"
+      />
+      <Select
+        label={t("extractFields.presets.label", "Preset template")}
+        value={matchPreset(parameters.fields)}
+        onChange={applyPreset}
+        data={[
+          ...FIELD_PRESET_IDS.map((preset) => ({
+            value: preset,
+            label: presetLabels[preset],
+          })),
+          {
+            value: CUSTOM_PRESET,
+            label: t("extractFields.presets.custom", "Custom"),
+          },
+        ]}
+        disabled={disabled}
+      />
+      <Text size="sm" fw={500}>
+        {t("extractFields.fields.label", "Fields to extract")}
+      </Text>
+      <div className={styles.schema}>
+        {parameters.fields.map((row, index) => (
+          <div key={index} className={styles.row}>
+            <TextInput
+              className={styles.nameInput}
+              aria-label={t("extractFields.fields.name", "Name")}
+              placeholder={t(
+                "extractFields.fields.namePlaceholder",
+                "invoice_number",
+              )}
+              value={row.name}
+              onChange={(event) =>
+                setRow(index, { name: event.currentTarget.value })
+              }
+              disabled={disabled}
+            />
+            <Select
+              className={styles.typeSelect}
+              aria-label={t("extractFields.fields.type", "Type")}
+              value={row.type}
+              onChange={(value) =>
+                setRow(index, { type: (value ?? "string") as FieldType })
+              }
+              data={FIELD_TYPES.map((type) => ({ value: type, label: type }))}
+              disabled={disabled}
+            />
+            <TextInput
+              className={styles.descriptionInput}
+              aria-label={t("extractFields.fields.description", "Description")}
+              placeholder={t(
+                "extractFields.fields.descriptionPlaceholder",
+                "What to look for",
+              )}
+              value={row.description}
+              onChange={(event) =>
+                setRow(index, { description: event.currentTarget.value })
+              }
+              disabled={disabled}
+            />
+            <Button
+              variant="quiet"
+              size="sm"
+              shape="circle"
+              leftSection={
+                <DeleteOutlineRoundedIcon style={{ fontSize: "1.1rem" }} />
+              }
+              aria-label={t("extractFields.fields.remove", "Remove field")}
+              onClick={() => removeRow(index)}
+              disabled={disabled}
+            />
+          </div>
+        ))}
+      </div>
+      <Group gap="xs">
+        <Button
+          variant="secondary"
+          size="sm"
+          leftSection={<AddRoundedIcon style={{ fontSize: "1rem" }} />}
+          onClick={() =>
+            onParameterChange("fields", [...parameters.fields, emptyFieldRow()])
+          }
+          disabled={disabled}
+        >
+          {t("extractFields.fields.add", "Add field")}
+        </Button>
+        <Button
+          variant="secondary"
+          size="sm"
+          leftSection={<AutoAwesomeRoundedIcon style={{ fontSize: "1rem" }} />}
+          onClick={suggestFields}
+          loading={suggesting}
+          disabled={disabled || !selectedFile}
+        >
+          {t("extractFields.suggest.button", "Suggest fields (AI)")}
+        </Button>
+      </Group>
+      {!selectedFile && (
+        <Text size="xs" c="dimmed">
+          {t(
+            "extractFields.suggest.needsFile",
+            "Select a file first to suggest fields",
+          )}
+        </Text>
+      )}
+      <Select
+        label={t("extractFields.mode.label", "Mode")}
+        value={parameters.mode}
+        onChange={(value) =>
+          onParameterChange("mode", (value ?? "auto") as DocparseMode)
+        }
+        data={[
+          { value: "auto", label: t("extractFields.mode.auto", "Auto") },
+          { value: "basic", label: t("extractFields.mode.basic", "Basic") },
+          {
+            value: "advanced",
+            label: t("extractFields.mode.advanced", "Advanced"),
+          },
+        ]}
+        disabled={disabled}
+      />
+      <Textarea
+        label={t("extractFields.instructions.label", "Instructions")}
+        placeholder={t(
+          "extractFields.instructions.placeholder",
+          "e.g. Amounts are in EUR unless stated otherwise",
+        )}
+        value={parameters.instructions}
+        onChange={(event) =>
+          onParameterChange("instructions", event.currentTarget.value)
+        }
+        minRows={2}
+        autosize
+        disabled={disabled}
+      />
+    </Stack>
+  );
+};
+
+export default ExtractFieldsSettings;

frontend/editor/src/proprietary/components/tools/docparse/FillTemplateSettings.tsx

@@ -0,0 +1,54 @@
+import { useTranslation } from "react-i18next";
+import { Stack, Text, Textarea } from "@mantine/core";
+import type { ToolAutomationSettingsProps } from "@app/hooks/tools/shared/toolOperationTypes";
+import {
+  isJsonObjectString,
+  type FillTemplateParameters,
+} from "@app/hooks/tools/fillTemplate/useFillTemplateParameters";
+import DocparseToolIntro from "@app/components/tools/docparse/DocparseToolIntro";
+
+const FillTemplateSettings = ({
+  parameters,
+  onParameterChange,
+  disabled,
+}: ToolAutomationSettingsProps<FillTemplateParameters>) => {
+  const { t } = useTranslation();
+
+  const dataJson = parameters.dataJson;
+  const jsonError =
+    dataJson.trim().length > 0 && !isJsonObjectString(dataJson)
+      ? t("fillTemplate.data.invalid", "Enter a valid JSON object")
+      : null;
+
+  return (
+    <Stack gap="sm">
+      <DocparseToolIntro
+        description={t(
+          "fillTemplate.intro",
+          "Replaces the placeholders in a Word (.docx) template with your JSON data and returns the filled document - deterministic, no AI involved.",
+        )}
+        showFallbackNote={false}
+      />
+      <Text size="sm" c="dimmed">
+        {t(
+          "fillTemplate.hint",
+          "The input file must be a .docx template (not a PDF). Each placeholder in the template is replaced with the matching JSON value.",
+        )}
+      </Text>
+      <Textarea
+        label={t("fillTemplate.data.label", "Data (JSON)")}
+        placeholder='{"customer": "ACME Corp", "total": "128.00"}'
+        value={dataJson}
+        onChange={(event) =>
+          onParameterChange("dataJson", event.currentTarget.value)
+        }
+        minRows={5}
+        autosize
+        error={jsonError}
+        disabled={disabled}
+      />
+    </Stack>
+  );
+};
+
+export default FillTemplateSettings;

frontend/editor/src/proprietary/components/tools/docparse/ParseDocumentSettings.tsx

@@ -0,0 +1,78 @@
+import { useTranslation } from "react-i18next";
+import { Checkbox, Select, Stack } from "@mantine/core";
+import type { ToolAutomationSettingsProps } from "@app/hooks/tools/shared/toolOperationTypes";
+import type {
+  DocparseMode,
+  ParseDocumentParameters,
+} from "@app/hooks/tools/parseDocument/useParseDocumentParameters";
+import DocparseToolIntro from "@app/components/tools/docparse/DocparseToolIntro";
+
+const ParseDocumentSettings = ({
+  parameters,
+  onParameterChange,
+  disabled,
+}: ToolAutomationSettingsProps<ParseDocumentParameters>) => {
+  const { t } = useTranslation();
+
+  return (
+    <Stack gap="sm">
+      <DocparseToolIntro
+        description={t(
+          "parseDocument.intro",
+          "Reads the document's layout - headings, paragraphs, tables - and turns it into clean structured JSON or Markdown you can feed to other systems.",
+        )}
+        aiBadge="layout"
+      />
+      <Select
+        label={t("parseDocument.mode.label", "Mode")}
+        value={parameters.mode}
+        onChange={(value) =>
+          onParameterChange("mode", (value ?? "auto") as DocparseMode)
+        }
+        data={[
+          { value: "auto", label: t("parseDocument.mode.auto", "Auto") },
+          { value: "basic", label: t("parseDocument.mode.basic", "Basic") },
+          {
+            value: "advanced",
+            label: t("parseDocument.mode.advanced", "Advanced"),
+          },
+        ]}
+        disabled={disabled}
+      />
+      <Select
+        label={t("parseDocument.outputFormat.label", "Output format")}
+        value={parameters.outputFormat}
+        onChange={(value) =>
+          onParameterChange(
+            "outputFormat",
+            (value ?? "json") as ParseDocumentParameters["outputFormat"],
+          )
+        }
+        data={[
+          {
+            value: "json",
+            label: t("parseDocument.outputFormat.json", "JSON"),
+          },
+          {
+            value: "markdown",
+            label: t("parseDocument.outputFormat.markdown", "Markdown"),
+          },
+        ]}
+        disabled={disabled}
+      />
+      <Checkbox
+        label={t(
+          "parseDocument.withOcr.label",
+          "Apply OCR to scanned pages (recommended)",
+        )}
+        checked={parameters.withOcr}
+        onChange={(event) =>
+          onParameterChange("withOcr", event.currentTarget.checked)
+        }
+        disabled={disabled}
+      />
+    </Stack>
+  );
+};
+
+export default ParseDocumentSettings;

frontend/editor/src/proprietary/components/tools/docparse/SmartSplitSettings.tsx

@@ -0,0 +1,51 @@
+import { useTranslation } from "react-i18next";
+import { NumberInput, Stack, Textarea } from "@mantine/core";
+import type { ToolAutomationSettingsProps } from "@app/hooks/tools/shared/toolOperationTypes";
+import type { SmartSplitParameters } from "@app/hooks/tools/smartSplit/useSmartSplitParameters";
+import DocparseToolIntro from "@app/components/tools/docparse/DocparseToolIntro";
+
+const SmartSplitSettings = ({
+  parameters,
+  onParameterChange,
+  disabled,
+}: ToolAutomationSettingsProps<SmartSplitParameters>) => {
+  const { t } = useTranslation();
+
+  return (
+    <Stack gap="sm">
+      <DocparseToolIntro
+        description={t(
+          "smartSplit.intro",
+          "Describe where sub-documents start in plain language and AI reads the content to find those boundaries - no page numbers needed.",
+        )}
+        aiBadge="llm"
+      />
+      <Textarea
+        label={t("smartSplit.rule.label", "Split rule")}
+        placeholder={t(
+          "smartSplit.rule.placeholder",
+          "e.g. Start a new document at every invoice header",
+        )}
+        value={parameters.rule}
+        onChange={(event) =>
+          onParameterChange("rule", event.currentTarget.value)
+        }
+        minRows={3}
+        autosize
+        disabled={disabled}
+      />
+      <NumberInput
+        label={t("smartSplit.maxParts.label", "Maximum parts")}
+        value={parameters.maxParts}
+        onChange={(value) =>
+          onParameterChange("maxParts", typeof value === "number" ? value : 1)
+        }
+        min={1}
+        max={100}
+        disabled={disabled}
+      />
+    </Stack>
+  );
+};
+
+export default SmartSplitSettings;

frontend/editor/src/proprietary/data/useProprietaryToolRegistry.tsx

@@ -0,0 +1,128 @@
+import { useMemo } from "react";
+import { useTranslation } from "react-i18next";
+
+import LocalIcon from "@app/components/shared/LocalIcon";
+import {
+  SubcategoryId,
+  ToolCategoryId,
+  type ProprietaryToolRegistry,
+} from "@app/data/toolsTaxonomy";
+import { asRegistryConfig } from "@app/hooks/tools/shared/toolOperationTypes";
+import { useDocparseEnabled } from "@app/hooks/useDocparseEnabled";
+import { parseDocumentOperationConfig } from "@app/hooks/tools/parseDocument/parseDocumentOperationConfig";
+import { extractFieldsOperationConfig } from "@app/hooks/tools/extractFields/extractFieldsOperationConfig";
+import { smartSplitOperationConfig } from "@app/hooks/tools/smartSplit/smartSplitOperationConfig";
+import { chunkDocumentOperationConfig } from "@app/hooks/tools/chunkDocument/chunkDocumentOperationConfig";
+import { fillTemplateOperationConfig } from "@app/hooks/tools/fillTemplate/fillTemplateOperationConfig";
+import ParseDocument from "@app/tools/ParseDocument";
+import ExtractFields from "@app/tools/ExtractFields";
+import SmartSplit from "@app/tools/SmartSplit";
+import ChunkDocument from "@app/tools/ChunkDocument";
+import FillTemplate from "@app/tools/FillTemplate";
+import { getSynonyms } from "@app/utils/toolSynonyms";
+
+const toolIcon = (icon: string) => (
+  <LocalIcon icon={icon} width="1.5rem" height="1.5rem" />
+);
+
+/**
+ * Proprietary tool registry extension - the DocParse tool family.
+ * Overrides the empty stub at {@code core/data/useProprietaryToolRegistry.tsx}.
+ * Hidden entirely while the backend reports docparse disabled.
+ */
+export function useProprietaryToolRegistry(): ProprietaryToolRegistry {
+  const { t } = useTranslation();
+  const docparseEnabled = useDocparseEnabled();
+
+  return useMemo(() => {
+    if (!docparseEnabled) return {} as ProprietaryToolRegistry;
+    return {
+      parseDocument: {
+        icon: toolIcon("quick-reference-all-outline-rounded"),
+        name: t("home.parseDocument.title", "Parse Document"),
+        component: ParseDocument,
+        description: t(
+          "home.parseDocument.desc",
+          "Layout-aware parsing to structured JSON or Markdown, with optional OCR",
+        ),
+        categoryId: ToolCategoryId.STANDARD_TOOLS,
+        subcategoryId: SubcategoryId.DOCUMENT_INTELLIGENCE,
+        maxFiles: 1,
+        endpoints: ["parse-document"],
+        operationConfig: asRegistryConfig(parseDocumentOperationConfig),
+        automationSettings: null,
+        synonyms: getSynonyms(t, "parseDocument"),
+        versionStatus: "beta",
+      },
+      extractFields: {
+        icon: toolIcon("fact-check-outline-rounded"),
+        name: t("home.extractFields.title", "Extract Fields"),
+        component: ExtractFields,
+        description: t(
+          "home.extractFields.desc",
+          "Pull typed fields out of a document with confidence scores and citations",
+        ),
+        categoryId: ToolCategoryId.STANDARD_TOOLS,
+        subcategoryId: SubcategoryId.DOCUMENT_INTELLIGENCE,
+        maxFiles: 1,
+        endpoints: ["extract-fields"],
+        operationConfig: asRegistryConfig(extractFieldsOperationConfig),
+        automationSettings: null,
+        synonyms: getSynonyms(t, "extractFields"),
+        versionStatus: "beta",
+      },
+      smartSplit: {
+        icon: toolIcon("content-cut-rounded"),
+        name: t("home.smartSplit.title", "Smart Split"),
+        component: SmartSplit,
+        description: t(
+          "home.smartSplit.desc",
+          "Split a PDF into sub-documents using a natural-language boundary rule",
+        ),
+        categoryId: ToolCategoryId.STANDARD_TOOLS,
+        subcategoryId: SubcategoryId.DOCUMENT_INTELLIGENCE,
+        maxFiles: 1,
+        endpoints: ["smart-split"],
+        operationConfig: asRegistryConfig(smartSplitOperationConfig),
+        automationSettings: null,
+        synonyms: getSynonyms(t, "smartSplit"),
+        versionStatus: "beta",
+      },
+      chunkDocument: {
+        icon: toolIcon("layers-outline-rounded"),
+        name: t("home.chunkDocument.title", "Prepare for RAG"),
+        component: ChunkDocument,
+        description: t(
+          "home.chunkDocument.desc",
+          "Layout-aware parse to structure-aware chunks with heading breadcrumbs and page ranges, ready to embed",
+        ),
+        categoryId: ToolCategoryId.STANDARD_TOOLS,
+        subcategoryId: SubcategoryId.DOCUMENT_INTELLIGENCE,
+        maxFiles: 1,
+        endpoints: ["chunk-document"],
+        operationConfig: asRegistryConfig(chunkDocumentOperationConfig),
+        automationSettings: null,
+        synonyms: getSynonyms(t, "chunkDocument"),
+        versionStatus: "beta",
+      },
+      fillTemplate: {
+        icon: toolIcon("assignment-outline-rounded"),
+        name: t("home.fillTemplate.title", "Fill Template"),
+        component: FillTemplate,
+        description: t(
+          "home.fillTemplate.desc",
+          "Fill a DOCX template's placeholders from JSON data",
+        ),
+        categoryId: ToolCategoryId.STANDARD_TOOLS,
+        subcategoryId: SubcategoryId.DOCUMENT_INTELLIGENCE,
+        maxFiles: 1,
+        supportedFormats: ["docx"],
+        endpoints: ["fill-template"],
+        operationConfig: asRegistryConfig(fillTemplateOperationConfig),
+        automationSettings: null,
+        synonyms: getSynonyms(t, "fillTemplate"),
+        versionStatus: "beta",
+      },
+    } as ProprietaryToolRegistry;
+  }, [t, docparseEnabled]);
+}

frontend/editor/src/proprietary/hooks/tools/chunkDocument/chunkDocumentOperationConfig.test.ts

@@ -0,0 +1,42 @@
+import { describe, expect, test } from "vitest";
+import {
+  buildChunkDocumentFormData,
+  chunksFromResponse,
+  chunksToJsonl,
+} from "@app/hooks/tools/chunkDocument/chunkDocumentOperationConfig";
+import { defaultParameters } from "@app/hooks/tools/chunkDocument/useChunkDocumentParameters";
+
+describe("chunkDocument operation helpers", () => {
+  test("accepts both bare-array and wrapped chunk responses", () => {
+    const chunks = [{ text: "a" }, { text: "b" }];
+    expect(chunksFromResponse(chunks)).toEqual(chunks);
+    expect(chunksFromResponse({ chunks })).toEqual(chunks);
+    expect(chunksFromResponse({ nope: true })).toEqual([]);
+    expect(chunksFromResponse(null)).toEqual([]);
+  });
+
+  test("emits one JSON document per JSONL line", () => {
+    const jsonl = chunksToJsonl([
+      { text: "a", page: 1 },
+      { text: "b", page: 2 },
+    ]);
+    const lines = jsonl.split("\n");
+    expect(lines).toHaveLength(2);
+    expect(lines.map((line) => JSON.parse(line))).toEqual([
+      { text: "a", page: 1 },
+      { text: "b", page: 2 },
+    ]);
+  });
+
+  test("builds the multipart form the backend contract expects", () => {
+    const file = new File(["pdf"], "doc.pdf", { type: "application/pdf" });
+    const form = buildChunkDocumentFormData(
+      { ...defaultParameters, chunkSize: 800, overlap: 50, mode: "advanced" },
+      file,
+    );
+    expect(form.get("fileInput")).toBe(file);
+    expect(form.get("chunkSize")).toBe("800");
+    expect(form.get("overlap")).toBe("50");
+    expect(form.get("mode")).toBe("advanced");
+  });
+});

frontend/editor/src/proprietary/hooks/tools/chunkDocument/chunkDocumentOperationConfig.ts

@@ -0,0 +1,64 @@
+import apiClient from "@app/services/apiClient";
+import {
+  defineCustomTool,
+  CustomProcessorResult,
+} from "@app/hooks/tools/shared/toolOperationTypes";
+import { deriveName } from "@app/hooks/tools/shared/docparseFilenames";
+import {
+  ChunkDocumentParameters,
+  defaultParameters,
+} from "@app/hooks/tools/chunkDocument/useChunkDocumentParameters";
+
+export const CHUNK_DOCUMENT_ENDPOINT = "/api/v1/docparse/chunk-document";
+
+export const buildChunkDocumentFormData = (
+  parameters: ChunkDocumentParameters,
+  file: File,
+): FormData => {
+  const formData = new FormData();
+  formData.append("fileInput", file);
+  formData.append("chunkSize", String(parameters.chunkSize));
+  formData.append("overlap", String(parameters.overlap));
+  formData.append("mode", parameters.mode);
+  return formData;
+};
+
+/** The chunks array, whether the backend returns it bare or wrapped. */
+export function chunksFromResponse(data: unknown): unknown[] {
+  if (Array.isArray(data)) return data;
+  const wrapped = (data as { chunks?: unknown[] } | null)?.chunks;
+  return Array.isArray(wrapped) ? wrapped : [];
+}
+
+/** JSON chunks -> one JSONL line per chunk, the standard RAG-ingest shape. */
+export function chunksToJsonl(chunks: unknown[]): string {
+  return chunks.map((chunk) => JSON.stringify(chunk)).join("\n");
+}
+
+const processChunkDocument = async (
+  parameters: ChunkDocumentParameters,
+  files: File[],
+): Promise<CustomProcessorResult> => {
+  if (files.length === 0) return { files: [] };
+
+  const [inputFile] = files;
+  const response = await apiClient.post<unknown>(
+    CHUNK_DOCUMENT_ENDPOINT,
+    buildChunkDocumentFormData(parameters, inputFile),
+  );
+
+  const resultFile = new File(
+    [chunksToJsonl(chunksFromResponse(response.data))],
+    deriveName(inputFile.name, ".chunks.jsonl"),
+    { type: "application/x-ndjson" },
+  );
+  return { files: [resultFile] };
+};
+
+export const chunkDocumentOperationConfig =
+  defineCustomTool<ChunkDocumentParameters>({
+    operationType: "chunkDocument",
+    endpoint: CHUNK_DOCUMENT_ENDPOINT,
+    customProcessor: processChunkDocument,
+    defaultParameters,
+  });

frontend/editor/src/proprietary/hooks/tools/chunkDocument/useChunkDocumentOperation.ts

@@ -0,0 +1,15 @@
+import { useTranslation } from "react-i18next";
+import { useToolOperation } from "@app/hooks/tools/shared/useToolOperation";
+import { createStandardErrorHandler } from "@app/utils/toolErrorHandler";
+import { chunkDocumentOperationConfig } from "@app/hooks/tools/chunkDocument/chunkDocumentOperationConfig";
+
+export const useChunkDocumentOperation = () => {
+  const { t } = useTranslation();
+
+  return useToolOperation({
+    ...chunkDocumentOperationConfig,
+    getErrorMessage: createStandardErrorHandler(
+      t("chunkDocument.error.failed", "Failed to chunk document"),
+    ),
+  });
+};

frontend/editor/src/proprietary/hooks/tools/chunkDocument/useChunkDocumentParameters.ts

@@ -0,0 +1,34 @@
+import { BaseParameters } from "@app/types/parameters";
+import {
+  useBaseParameters,
+  BaseParametersHook,
+} from "@app/hooks/tools/shared/useBaseParameters";
+import type { DocparseMode } from "@app/hooks/tools/parseDocument/useParseDocumentParameters";
+
+export interface ChunkDocumentParameters extends BaseParameters {
+  /** Target chunk size in characters. */
+  chunkSize: number;
+  /** Characters of overlap carried between neighbouring chunks. */
+  overlap: number;
+  mode: DocparseMode;
+}
+
+export const defaultParameters: ChunkDocumentParameters = {
+  chunkSize: 1000,
+  overlap: 100,
+  mode: "auto",
+};
+
+export type ChunkDocumentParametersHook =
+  BaseParametersHook<ChunkDocumentParameters>;
+
+export const useChunkDocumentParameters = (): ChunkDocumentParametersHook => {
+  return useBaseParameters({
+    defaultParameters,
+    endpointName: "chunk-document",
+    validateFn: (params) =>
+      params.chunkSize > 0 &&
+      params.overlap >= 0 &&
+      params.overlap < params.chunkSize,
+  });
+};

frontend/editor/src/proprietary/hooks/tools/extractFields/extractFieldsOperationConfig.ts

@@ -0,0 +1,71 @@
+import apiClient from "@app/services/apiClient";
+import {
+  defineCustomTool,
+  CustomProcessorResult,
+} from "@app/hooks/tools/shared/toolOperationTypes";
+import { deriveName } from "@app/hooks/tools/shared/docparseFilenames";
+import { rowsToSchemaString } from "@app/hooks/tools/extractFields/fieldsSchema";
+import {
+  ExtractFieldsParameters,
+  defaultParameters,
+} from "@app/hooks/tools/extractFields/useExtractFieldsParameters";
+
+// The JSON variant; the plain /extract-fields form is the pipeline shape.
+export const EXTRACT_FIELDS_ENDPOINT = "/api/v1/docparse/extract-fields/json";
+
+/** One extracted field in the backend's extraction report. */
+export interface ExtractedField {
+  name: string;
+  value: unknown;
+  confidence: number;
+  citations?: { page: number; bbox?: number[] | null; quote?: string }[];
+}
+
+export interface ExtractFieldsResult {
+  mode: string;
+  fields: ExtractedField[];
+  overallConfidence?: number;
+}
+
+export const buildExtractFieldsFormData = (
+  parameters: ExtractFieldsParameters,
+  file: File,
+): FormData => {
+  const formData = new FormData();
+  formData.append("fileInput", file);
+  formData.append("fieldsSchema", rowsToSchemaString(parameters.fields));
+  formData.append("mode", parameters.mode);
+  if (parameters.instructions.trim()) {
+    formData.append("instructions", parameters.instructions.trim());
+  }
+  return formData;
+};
+
+/** POST the PDF + schema; keep the extraction report as a JSON result file. */
+const processExtractFields = async (
+  parameters: ExtractFieldsParameters,
+  files: File[],
+): Promise<CustomProcessorResult> => {
+  if (files.length === 0) return { files: [] };
+
+  const [inputFile] = files;
+  const response = await apiClient.post<ExtractFieldsResult>(
+    EXTRACT_FIELDS_ENDPOINT,
+    buildExtractFieldsFormData(parameters, inputFile),
+  );
+
+  const resultFile = new File(
+    [JSON.stringify(response.data, null, 2)],
+    deriveName(inputFile.name, ".fields.json"),
+    { type: "application/json" },
+  );
+  return { files: [resultFile] };
+};
+
+export const extractFieldsOperationConfig =
+  defineCustomTool<ExtractFieldsParameters>({
+    operationType: "extractFields",
+    endpoint: EXTRACT_FIELDS_ENDPOINT,
+    customProcessor: processExtractFields,
+    defaultParameters,
+  });

frontend/editor/src/proprietary/hooks/tools/extractFields/fieldsSchema.test.ts

@@ -0,0 +1,59 @@
+import { describe, expect, test } from "vitest";
+import {
+  emptyFieldRow,
+  namedRows,
+  rowsFromSchemaString,
+  rowsToSchemaString,
+  type FieldRow,
+} from "@app/hooks/tools/extractFields/fieldsSchema";
+
+const rows: FieldRow[] = [
+  { name: "invoice_number", type: "string", description: "The invoice id" },
+  { name: "total", type: "number", description: "" },
+  { name: "paid", type: "boolean", description: "Whether settled" },
+];
+
+describe("fieldsSchema", () => {
+  test("serializes rows to a JSON Schema object string", () => {
+    const schema = JSON.parse(rowsToSchemaString(rows));
+    expect(schema.type).toBe("object");
+    expect(schema.required).toEqual(["invoice_number", "total", "paid"]);
+    expect(schema.properties.invoice_number).toEqual({
+      type: "string",
+      description: "The invoice id",
+    });
+    // Blank descriptions are omitted, not sent as empty strings.
+    expect(schema.properties.total).toEqual({ type: "number" });
+  });
+
+  test("round-trips rows through the schema string", () => {
+    expect(rowsFromSchemaString(rowsToSchemaString(rows))).toEqual(rows);
+  });
+
+  test("ignores unnamed builder rows", () => {
+    const withBlank = [...rows, emptyFieldRow()];
+    expect(namedRows(withBlank)).toHaveLength(3);
+    const schema = JSON.parse(rowsToSchemaString(withBlank));
+    expect(Object.keys(schema.properties)).toHaveLength(3);
+  });
+
+  test("trims names and descriptions on serialize", () => {
+    const schema = JSON.parse(
+      rowsToSchemaString([
+        { name: "  due_date  ", type: "string", description: " When due " },
+      ]),
+    );
+    expect(schema.properties.due_date).toEqual({
+      type: "string",
+      description: "When due",
+    });
+  });
+
+  test("parses junk defensively", () => {
+    expect(rowsFromSchemaString("not json")).toEqual([]);
+    expect(rowsFromSchemaString("{}")).toEqual([]);
+    expect(
+      rowsFromSchemaString('{"properties":{"x":{"type":"weird"}}}'),
+    ).toEqual([{ name: "x", type: "string", description: "" }]);
+  });
+});

frontend/editor/src/proprietary/hooks/tools/extractFields/fieldsSchema.ts

@@ -0,0 +1,65 @@
+/**
+ * The Extract Fields schema builder model: flat rows of {name, type, description}
+ * serialized to the JSON Schema object string the backend's fieldsSchema expects.
+ */
+
+export const FIELD_TYPES = ["string", "number", "integer", "boolean"] as const;
+
+export type FieldType = (typeof FIELD_TYPES)[number];
+
+export interface FieldRow {
+  name: string;
+  type: FieldType;
+  description: string;
+}
+
+export const emptyFieldRow = (): FieldRow => ({
+  name: "",
+  type: "string",
+  description: "",
+});
+
+/** Rows with a non-empty name, i.e. the ones worth serializing. */
+export function namedRows(rows: FieldRow[]): FieldRow[] {
+  return rows.filter((row) => row.name.trim().length > 0);
+}
+
+/** Serialize builder rows to the JSON Schema string sent as fieldsSchema. */
+export function rowsToSchemaString(rows: FieldRow[]): string {
+  const properties: Record<string, { type: FieldType; description?: string }> =
+    {};
+  for (const row of namedRows(rows)) {
+    properties[row.name.trim()] = {
+      type: row.type,
+      ...(row.description.trim()
+        ? { description: row.description.trim() }
+        : {}),
+    };
+  }
+  return JSON.stringify({
+    type: "object",
+    properties,
+    required: Object.keys(properties),
+  });
+}
+
+const isFieldType = (value: unknown): value is FieldType =>
+  typeof value === "string" && FIELD_TYPES.includes(value as FieldType);
+
+/** Parse a fieldsSchema string back into builder rows; [] on anything unusable. */
+export function rowsFromSchemaString(schema: string): FieldRow[] {
+  try {
+    const parsed = JSON.parse(schema) as {
+      properties?: Record<string, { type?: unknown; description?: unknown }>;
+    };
+    if (!parsed || typeof parsed !== "object" || !parsed.properties) return [];
+    return Object.entries(parsed.properties).map(([name, prop]) => ({
+      name,
+      type: isFieldType(prop?.type) ? prop.type : "string",
+      description:
+        typeof prop?.description === "string" ? prop.description : "",
+    }));
+  } catch {
+    return [];
+  }
+}

frontend/editor/src/proprietary/hooks/tools/extractFields/fieldsSchemaPresets.test.ts

@@ -0,0 +1,47 @@
+import { describe, expect, test } from "vitest";
+import {
+  emptyFieldRow,
+  rowsFromSchemaString,
+  rowsToSchemaString,
+} from "@app/hooks/tools/extractFields/fieldsSchema";
+import {
+  CUSTOM_PRESET,
+  FIELD_PRESET_IDS,
+  matchPreset,
+  presetRows,
+} from "@app/hooks/tools/extractFields/fieldsSchemaPresets";
+
+describe("fieldsSchemaPresets", () => {
+  test("every preset has 4-6 named, described fields", () => {
+    for (const preset of FIELD_PRESET_IDS) {
+      const rows = presetRows(preset);
+      expect(rows.length).toBeGreaterThanOrEqual(4);
+      expect(rows.length).toBeLessThanOrEqual(6);
+      for (const row of rows) {
+        expect(row.name).toMatch(/^[a-z0-9_]+$/);
+        expect(row.description.length).toBeGreaterThan(0);
+      }
+    }
+  });
+
+  test("preset rows survive the schema-string round trip", () => {
+    for (const preset of FIELD_PRESET_IDS) {
+      const rows = presetRows(preset);
+      expect(rowsFromSchemaString(rowsToSchemaString(rows))).toEqual(rows);
+    }
+  });
+
+  test("matchPreset spots untouched presets and demotes edits to custom", () => {
+    const rows = presetRows("invoice");
+    expect(matchPreset(rows)).toBe("invoice");
+    rows[0] = { ...rows[0], name: "order_ref" };
+    expect(matchPreset(rows)).toBe(CUSTOM_PRESET);
+    expect(matchPreset([emptyFieldRow()])).toBe(CUSTOM_PRESET);
+  });
+
+  test("presetRows hands out fresh copies, not shared references", () => {
+    const first = presetRows("receipt");
+    first[0].name = "mutated";
+    expect(presetRows("receipt")[0].name).not.toBe("mutated");
+  });
+});

frontend/editor/src/proprietary/hooks/tools/extractFields/fieldsSchemaPresets.ts

@@ -0,0 +1,151 @@
+import type { FieldRow } from "@app/hooks/tools/extractFields/fieldsSchema";
+
+/**
+ * Ready-made extraction schemas for the common document types, so users start
+ * from a sensible field list instead of a blank builder.
+ */
+
+export const FIELD_PRESET_IDS = [
+  "invoice",
+  "receipt",
+  "contract",
+  "purchaseOrder",
+] as const;
+
+export type FieldPresetId = (typeof FIELD_PRESET_IDS)[number];
+
+/** The select value when the rows match no preset (hand-built schema). */
+export const CUSTOM_PRESET = "custom" as const;
+
+const PRESETS: Record<FieldPresetId, FieldRow[]> = {
+  invoice: [
+    {
+      name: "invoice_number",
+      type: "string",
+      description: "The invoice identifier",
+    },
+    {
+      name: "invoice_date",
+      type: "string",
+      description: "Date the invoice was issued (ISO format if possible)",
+    },
+    {
+      name: "vendor_name",
+      type: "string",
+      description: "Name of the company issuing the invoice",
+    },
+    {
+      name: "total_amount",
+      type: "number",
+      description: "Grand total including tax",
+    },
+    {
+      name: "currency",
+      type: "string",
+      description: "Currency code or symbol",
+    },
+    {
+      name: "due_date",
+      type: "string",
+      description: "Date payment is due",
+    },
+  ],
+  receipt: [
+    {
+      name: "merchant_name",
+      type: "string",
+      description: "Store or merchant name",
+    },
+    {
+      name: "purchase_date",
+      type: "string",
+      description: "Date of the purchase",
+    },
+    { name: "total_amount", type: "number", description: "Total paid" },
+    {
+      name: "tax_amount",
+      type: "number",
+      description: "Tax portion of the total",
+    },
+    {
+      name: "payment_method",
+      type: "string",
+      description: "How it was paid, e.g. card or cash",
+    },
+  ],
+  contract: [
+    {
+      name: "party_a",
+      type: "string",
+      description: "First contracting party's legal name",
+    },
+    {
+      name: "party_b",
+      type: "string",
+      description: "Second contracting party's legal name",
+    },
+    {
+      name: "effective_date",
+      type: "string",
+      description: "Date the agreement takes effect",
+    },
+    {
+      name: "termination_date",
+      type: "string",
+      description: "Date the agreement ends or renews",
+    },
+    {
+      name: "governing_law",
+      type: "string",
+      description: "Jurisdiction governing the agreement",
+    },
+    {
+      name: "auto_renews",
+      type: "boolean",
+      description: "Whether the contract renews automatically",
+    },
+  ],
+  purchaseOrder: [
+    {
+      name: "po_number",
+      type: "string",
+      description: "Purchase order identifier",
+    },
+    { name: "order_date", type: "string", description: "Date of the order" },
+    {
+      name: "supplier_name",
+      type: "string",
+      description: "Supplier the order is placed with",
+    },
+    {
+      name: "delivery_date",
+      type: "string",
+      description: "Requested or promised delivery date",
+    },
+    { name: "total_amount", type: "number", description: "Order total" },
+  ],
+};
+
+/** A fresh copy of the preset's rows, safe to mutate in the builder. */
+export function presetRows(preset: FieldPresetId): FieldRow[] {
+  return PRESETS[preset].map((row) => ({ ...row }));
+}
+
+/** The preset the rows exactly match, or "custom" for anything hand-edited. */
+export function matchPreset(rows: FieldRow[]): FieldPresetId | "custom" {
+  for (const preset of FIELD_PRESET_IDS) {
+    const candidate = PRESETS[preset];
+    if (
+      rows.length === candidate.length &&
+      rows.every(
+        (row, i) =>
+          row.name === candidate[i].name &&
+          row.type === candidate[i].type &&
+          row.description === candidate[i].description,
+      )
+    ) {
+      return preset;
+    }
+  }
+  return CUSTOM_PRESET;
+}

frontend/editor/src/proprietary/hooks/tools/extractFields/suggestSchema.test.ts

@@ -0,0 +1,35 @@
+import { describe, expect, test } from "vitest";
+import { suggestedFieldsToRows } from "@app/hooks/tools/extractFields/suggestSchema";
+
+describe("suggestedFieldsToRows", () => {
+  test("maps engine proposals to builder rows", () => {
+    expect(
+      suggestedFieldsToRows([
+        { name: "invoice_number", type: "string", description: "The id" },
+        { name: "total", type: "number" },
+      ]),
+    ).toEqual([
+      { name: "invoice_number", type: "string", description: "The id" },
+      { name: "total", type: "number", description: "" },
+    ]);
+  });
+
+  test("coerces unknown types to string and trims values", () => {
+    expect(
+      suggestedFieldsToRows([
+        { name: "  due_date ", type: "date", description: " When due " },
+      ]),
+    ).toEqual([{ name: "due_date", type: "string", description: "When due" }]);
+  });
+
+  test("drops unusable entries defensively", () => {
+    expect(
+      suggestedFieldsToRows([
+        { name: "", type: "string" },
+        { type: "string", description: "no name" },
+        undefined as never,
+      ]),
+    ).toEqual([]);
+    expect(suggestedFieldsToRows(undefined)).toEqual([]);
+  });
+});

frontend/editor/src/proprietary/hooks/tools/extractFields/suggestSchema.ts

@@ -0,0 +1,53 @@
+import apiClient from "@app/services/apiClient";
+import {
+  FIELD_TYPES,
+  type FieldRow,
+  type FieldType,
+} from "@app/hooks/tools/extractFields/fieldsSchema";
+
+export const SUGGEST_SCHEMA_ENDPOINT = "/api/v1/docparse/suggest-schema";
+
+/** Cap on proposed fields; keeps the builder scannable in the side panel. */
+export const SUGGEST_MAX_FIELDS = 8;
+
+/** One field the backend proposes for an extraction schema. */
+export interface SuggestedSchemaField {
+  name?: string;
+  type?: string;
+  description?: string;
+}
+
+export interface SuggestSchemaResponse {
+  mode?: string;
+  fields?: SuggestedSchemaField[];
+}
+
+const isFieldType = (value: unknown): value is FieldType =>
+  typeof value === "string" && FIELD_TYPES.includes(value as FieldType);
+
+/** Map the engine's proposals to builder rows, dropping unusable entries. */
+export function suggestedFieldsToRows(
+  fields: SuggestedSchemaField[] | undefined,
+): FieldRow[] {
+  if (!Array.isArray(fields)) return [];
+  return fields
+    .filter((field) => typeof field?.name === "string" && field.name.trim())
+    .map((field) => ({
+      name: field.name!.trim(),
+      type: isFieldType(field.type) ? field.type : "string",
+      description:
+        typeof field.description === "string" ? field.description.trim() : "",
+    }));
+}
+
+/** POST the document; get AI-proposed schema rows for the builder. */
+export async function requestSuggestedFields(file: File): Promise<FieldRow[]> {
+  const formData = new FormData();
+  formData.append("fileInput", file);
+  formData.append("maxFields", String(SUGGEST_MAX_FIELDS));
+  const response = await apiClient.post<SuggestSchemaResponse>(
+    SUGGEST_SCHEMA_ENDPOINT,
+    formData,
+  );
+  return suggestedFieldsToRows(response.data?.fields);
+}

frontend/editor/src/proprietary/hooks/tools/extractFields/useExtractFieldsOperation.ts

@@ -0,0 +1,15 @@
+import { useTranslation } from "react-i18next";
+import { useToolOperation } from "@app/hooks/tools/shared/useToolOperation";
+import { createStandardErrorHandler } from "@app/utils/toolErrorHandler";
+import { extractFieldsOperationConfig } from "@app/hooks/tools/extractFields/extractFieldsOperationConfig";
+
+export const useExtractFieldsOperation = () => {
+  const { t } = useTranslation();
+
+  return useToolOperation({
+    ...extractFieldsOperationConfig,
+    getErrorMessage: createStandardErrorHandler(
+      t("extractFields.error.failed", "Failed to extract fields"),
+    ),
+  });
+};

frontend/editor/src/proprietary/hooks/tools/extractFields/useExtractFieldsParameters.ts

@@ -0,0 +1,35 @@
+import { BaseParameters } from "@app/types/parameters";
+import {
+  useBaseParameters,
+  BaseParametersHook,
+} from "@app/hooks/tools/shared/useBaseParameters";
+import type { DocparseMode } from "@app/hooks/tools/shared/docparseTypes";
+import {
+  emptyFieldRow,
+  namedRows,
+  type FieldRow,
+} from "@app/hooks/tools/extractFields/fieldsSchema";
+
+export interface ExtractFieldsParameters extends BaseParameters {
+  /** Schema-builder rows; serialized to fieldsSchema on execute. */
+  fields: FieldRow[];
+  instructions: string;
+  mode: DocparseMode;
+}
+
+export const defaultParameters: ExtractFieldsParameters = {
+  fields: [emptyFieldRow()],
+  instructions: "",
+  mode: "auto",
+};
+
+export type ExtractFieldsParametersHook =
+  BaseParametersHook<ExtractFieldsParameters>;
+
+export const useExtractFieldsParameters = (): ExtractFieldsParametersHook => {
+  return useBaseParameters({
+    defaultParameters,
+    endpointName: "extract-fields",
+    validateFn: (params) => namedRows(params.fields).length > 0,
+  });
+};

frontend/editor/src/proprietary/hooks/tools/fillTemplate/fillTemplateOperationConfig.ts

@@ -0,0 +1,55 @@
+import apiClient from "@app/services/apiClient";
+import {
+  defineCustomTool,
+  CustomProcessorResult,
+} from "@app/hooks/tools/shared/toolOperationTypes";
+import { deriveName } from "@app/hooks/tools/shared/docparseFilenames";
+import {
+  FillTemplateParameters,
+  defaultParameters,
+} from "@app/hooks/tools/fillTemplate/useFillTemplateParameters";
+
+export const FILL_TEMPLATE_ENDPOINT = "/api/v1/docparse/fill-template";
+
+const DOCX_TYPE =
+  "application/vnd.openxmlformats-officedocument.wordprocessingml.document";
+
+export const buildFillTemplateFormData = (
+  parameters: FillTemplateParameters,
+  file: File,
+): FormData => {
+  const formData = new FormData();
+  formData.append("templateFile", file);
+  formData.append("data", parameters.dataJson.trim());
+  return formData;
+};
+
+/** POST the DOCX template + data; the filled DOCX comes straight back. */
+const processFillTemplate = async (
+  parameters: FillTemplateParameters,
+  files: File[],
+): Promise<CustomProcessorResult> => {
+  if (files.length === 0) return { files: [] };
+
+  const [templateFile] = files;
+  const response = await apiClient.post<Blob>(
+    FILL_TEMPLATE_ENDPOINT,
+    buildFillTemplateFormData(parameters, templateFile),
+    { responseType: "blob" },
+  );
+
+  const resultFile = new File(
+    [response.data],
+    deriveName(templateFile.name, "-filled.docx"),
+    { type: DOCX_TYPE },
+  );
+  return { files: [resultFile] };
+};
+
+export const fillTemplateOperationConfig =
+  defineCustomTool<FillTemplateParameters>({
+    operationType: "fillTemplate",
+    endpoint: FILL_TEMPLATE_ENDPOINT,
+    customProcessor: processFillTemplate,
+    defaultParameters,
+  });

frontend/editor/src/proprietary/hooks/tools/fillTemplate/useFillTemplateOperation.ts

@@ -0,0 +1,15 @@
+import { useTranslation } from "react-i18next";
+import { useToolOperation } from "@app/hooks/tools/shared/useToolOperation";
+import { createStandardErrorHandler } from "@app/utils/toolErrorHandler";
+import { fillTemplateOperationConfig } from "@app/hooks/tools/fillTemplate/fillTemplateOperationConfig";
+
+export const useFillTemplateOperation = () => {
+  const { t } = useTranslation();
+
+  return useToolOperation({
+    ...fillTemplateOperationConfig,
+    getErrorMessage: createStandardErrorHandler(
+      t("fillTemplate.error.failed", "Failed to fill template"),
+    ),
+  });
+};

frontend/editor/src/proprietary/hooks/tools/fillTemplate/useFillTemplateParameters.ts

@@ -0,0 +1,37 @@
+import { BaseParameters } from "@app/types/parameters";
+import {
+  useBaseParameters,
+  BaseParametersHook,
+} from "@app/hooks/tools/shared/useBaseParameters";
+
+export interface FillTemplateParameters extends BaseParameters {
+  /** JSON object whose keys fill the template's placeholders. */
+  dataJson: string;
+}
+
+export const defaultParameters: FillTemplateParameters = {
+  dataJson: "",
+};
+
+/** True when the text parses to a plain JSON object. */
+export function isJsonObjectString(text: string): boolean {
+  try {
+    const parsed = JSON.parse(text);
+    return (
+      typeof parsed === "object" && parsed !== null && !Array.isArray(parsed)
+    );
+  } catch {
+    return false;
+  }
+}
+
+export type FillTemplateParametersHook =
+  BaseParametersHook<FillTemplateParameters>;
+
+export const useFillTemplateParameters = (): FillTemplateParametersHook => {
+  return useBaseParameters({
+    defaultParameters,
+    endpointName: "fill-template",
+    validateFn: (params) => isJsonObjectString(params.dataJson),
+  });
+};

frontend/editor/src/proprietary/hooks/tools/parseDocument/parseDocumentOperationConfig.ts

@@ -0,0 +1,56 @@
+import apiClient from "@app/services/apiClient";
+import {
+  defineCustomTool,
+  CustomProcessorResult,
+} from "@app/hooks/tools/shared/toolOperationTypes";
+import { deriveName } from "@app/hooks/tools/shared/docparseFilenames";
+import {
+  ParseDocumentParameters,
+  defaultParameters,
+} from "@app/hooks/tools/parseDocument/useParseDocumentParameters";
+
+// Not part of the generated ToolEndpoint union; DocParse is an optional addon.
+export const PARSE_DOCUMENT_ENDPOINT = "/api/v1/docparse/parse-document";
+
+export const buildParseDocumentFormData = (
+  parameters: ParseDocumentParameters,
+  file: File,
+): FormData => {
+  const formData = new FormData();
+  formData.append("fileInput", file);
+  formData.append("mode", parameters.mode);
+  formData.append("withOcr", String(parameters.withOcr));
+  formData.append("outputFormat", parameters.outputFormat);
+  return formData;
+};
+
+/** POST the PDF; wrap the JSON or markdown result as a downloadable file. */
+const processParseDocument = async (
+  parameters: ParseDocumentParameters,
+  files: File[],
+): Promise<CustomProcessorResult> => {
+  if (files.length === 0) return { files: [] };
+
+  const [inputFile] = files;
+  const response = await apiClient.post<Blob>(
+    PARSE_DOCUMENT_ENDPOINT,
+    buildParseDocumentFormData(parameters, inputFile),
+    { responseType: "blob" },
+  );
+
+  const isMarkdown = parameters.outputFormat === "markdown";
+  const resultFile = new File(
+    [response.data],
+    deriveName(inputFile.name, isMarkdown ? ".md" : ".parsed.json"),
+    { type: isMarkdown ? "text/markdown" : "application/json" },
+  );
+  return { files: [resultFile] };
+};
+
+export const parseDocumentOperationConfig =
+  defineCustomTool<ParseDocumentParameters>({
+    operationType: "parseDocument",
+    endpoint: PARSE_DOCUMENT_ENDPOINT,
+    customProcessor: processParseDocument,
+    defaultParameters,
+  });

frontend/editor/src/proprietary/hooks/tools/parseDocument/useParseDocumentOperation.ts

@@ -0,0 +1,15 @@
+import { useTranslation } from "react-i18next";
+import { useToolOperation } from "@app/hooks/tools/shared/useToolOperation";
+import { createStandardErrorHandler } from "@app/utils/toolErrorHandler";
+import { parseDocumentOperationConfig } from "@app/hooks/tools/parseDocument/parseDocumentOperationConfig";
+
+export const useParseDocumentOperation = () => {
+  const { t } = useTranslation();
+
+  return useToolOperation({
+    ...parseDocumentOperationConfig,
+    getErrorMessage: createStandardErrorHandler(
+      t("parseDocument.error.failed", "Failed to parse document"),
+    ),
+  });
+};

frontend/editor/src/proprietary/hooks/tools/parseDocument/useParseDocumentParameters.ts

@@ -0,0 +1,30 @@
+import { BaseParameters } from "@app/types/parameters";
+import {
+  useBaseParameters,
+  BaseParametersHook,
+} from "@app/hooks/tools/shared/useBaseParameters";
+
+export type DocparseMode = "auto" | "basic" | "advanced";
+
+export interface ParseDocumentParameters extends BaseParameters {
+  mode: DocparseMode;
+  outputFormat: "json" | "markdown";
+  withOcr: boolean;
+}
+
+// withOcr defaults true to match the API's default behaviour.
+export const defaultParameters: ParseDocumentParameters = {
+  mode: "auto",
+  outputFormat: "json",
+  withOcr: true,
+};
+
+export type ParseDocumentParametersHook =
+  BaseParametersHook<ParseDocumentParameters>;
+
+export const useParseDocumentParameters = (): ParseDocumentParametersHook => {
+  return useBaseParameters({
+    defaultParameters,
+    endpointName: "parse-document",
+  });
+};

frontend/editor/src/proprietary/hooks/tools/shared/docparseFilenames.ts

@@ -0,0 +1,12 @@
+/** Filename helpers shared by the DocParse tool processors. */
+
+/** "invoice.pdf" -> "invoice"; keeps names without an extension intact. */
+export function stripExtension(fileName: string): string {
+  const dot = fileName.lastIndexOf(".");
+  return dot > 0 ? fileName.slice(0, dot) : fileName;
+}
+
+/** Derived output name, e.g. deriveName("a.pdf", ".fields.json") -> "a.fields.json". */
+export function deriveName(inputName: string, suffix: string): string {
+  return `${stripExtension(inputName)}${suffix}`;
+}