@wildo-ai/saas-backend-lib 1.1.1 → 1.1.3

package/dist/esm/files/services/file-text-extraction.backend.service.js

@@ -50,8 +50,30 @@
  * owns that decision and documents the full precedence, including the branch that protects an
  * existing RAG corpus from a provider outage. Plain text never touches a provider.
  *
+ * ## Office / OpenDocument — six served LOCALLY, three provider-only (#434, closed 2026-09-09)
+ *
+ * The framework's `Documents` media group admits Word, Excel, PowerPoint and the OpenDocument
+ * trio, and until 2026-09-08 the extractor supported NONE of them: a corpus field using the group
+ * accepted uploads that could never contribute a chunk under any configuration, skipped at debug.
+ *
+ * The nine types are THREE technical problems, and reading them as one is what produced this
+ * header's earlier claim that a local library would cost "three to six" dependencies:
+ *
+ * - **OOXML** (`.docx` `.xlsx` `.pptx`) and **OpenDocument** (`.odt` `.ods` `.odp`) are ZIP+XML —
+ *   the SAME technique, so ONE module reads all six, on two dependencies rather than six
+ *   (`fflate` to unzip, `fast-xml-parser` to read). See
+ *   `office-document-text-extraction.backend.utils.ts`.
+ * - **Legacy binary Office** (`.doc` `.xls` `.ppt`) are OLE2 compound files from a different era.
+ *   A reader for them is a project, not a dependency, so they stay PROVIDER-ONLY.
+ *
+ * So the six take exactly the ladder PDF already takes, with the office extractor in the rung
+ * `pdfjs-dist` occupies: a configured provider is preferred when the field opted into it, the
+ * local library serves everything else, and a deployment with NO provider can finally read the
+ * documents its media group accepts. The three remaining types keep the conditional support and
+ * the precise refusal, because for them the provider genuinely is the only path.
+ *
  * Future:
- * - Office documents (docx via mammoth, xlsx/pptx via dedicated libs)
+ * - Legacy binary Office (`.doc` / `.xls` / `.ppt`) locally, if an OLE2 reader ever earns its place
  * - Audio/video transcript extraction
  *
  * The service operates against the storage provider contract — it downloads
@@ -74,6 +96,7 @@ import { injectable, inject, optional } from 'inversify';
 import { createRequire } from 'node:module';
 import { dirname, join } from 'node:path';
 import { ErrorType, ErrorCustomMessageReference } from '@wildo-ai/saas-models';
+import { extractOfficeDocumentText, LEGACY_BINARY_OFFICE_MIMES, LOCALLY_EXTRACTABLE_DOCUMENT_MIMES, } from '../../files/utils/office-document-text-extraction.backend.utils.js';
 import { SAAS_SERVICE_TYPES } from '../../inject/models/service-types.js';
 import { ErrorBuilderBackendService } from '../../errors/services/error-builder.backend.service.js';
 import { FileStorageRouterBackendService } from '../../files/storage/file-storage-router.backend.service.js';
@@ -94,6 +117,23 @@ const TEXT_MIME_EXACT = new Set([
     'application/toml',
 ]);
 const PDF_MIME = 'application/pdf';
+/**
+ * The three types that STILL need a provider — pre-2007 binary Office (`.doc` / `.xls` / `.ppt`).
+ *
+ * These are OLE2 compound files, not ZIP archives, so the local ZIP+XML extractor that serves the
+ * other six cannot read them and a reader for them is a project rather than a dependency. Support
+ * is therefore CONDITIONAL, and {@link isExtractionSupported} says so by ASKING the provider rather
+ * than asserting a capability the deployment may not have.
+ *
+ * Why conditional rather than unconditional-and-fail-later: RAG ingestion treats an unsupported
+ * MIME as "no corpus from this file", logged at DEBUG and never an error. Declaring these supported
+ * everywhere would turn a configuration ABSENCE into an ERROR log on every sweep run, for every
+ * application without an extraction provider — noise that says nothing an operator can act on.
+ *
+ * The six ZIP+XML types are no longer here, and that is the whole of #434's second half: they are
+ * served by {@link LOCALLY_EXTRACTABLE_DOCUMENT_MIMES} with no provider required.
+ */
+const PROVIDER_ONLY_DOCUMENT_MIMES = LEGACY_BINARY_OFFICE_MIMES;
 const moduleRequire = createRequire(import.meta.url);
 /**
  * Absolute path to `pdfjs-dist`'s bundled standard-font data, resolved from the INSTALLED package
@@ -167,7 +207,9 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
         const downloadResult = await provider.download({ fileId: file.fileId, storageProvider: file.storageProvider, storageRef: file.storageRef }, executionContext);
         try {
             try {
-                if (mimeType === PDF_MIME) {
+                if (mimeType === PDF_MIME
+                    || LOCALLY_EXTRACTABLE_DOCUMENT_MIMES.has(mimeType)
+                    || PROVIDER_ONLY_DOCUMENT_MIMES.has(mimeType)) {
                     return await this.extractFromDocument(downloadResult.stream, file.fileId, mimeType, options);
                 }
                 return await this.extractFromText(downloadResult.stream, file.fileId);
@@ -225,7 +267,46 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
             return true;
         if (TEXT_MIME_EXACT.has(lower))
             return true;
-        return TEXT_MIME_PREFIXES.some((prefix) => lower.startsWith(prefix));
+        if (TEXT_MIME_PREFIXES.some((prefix) => lower.startsWith(prefix)))
+            return true;
+        /*
+         * OOXML + OpenDocument: UNCONDITIONALLY supported (#434). A local ZIP+XML extractor serves all
+         * six, so the answer no longer depends on a deployment's provider selection — which is the
+         * point of the local library. A provider, where one is configured and the field opted in, is
+         * still preferred inside `extractFromDocument`; it is no longer the only path.
+         */
+        if (LOCALLY_EXTRACTABLE_DOCUMENT_MIMES.has(lower))
+            return true;
+        /*
+         * Legacy binary Office: supported IF AND ONLY IF a configured provider declares the type. Byte
+         * length is unknown at this question — it is asked before the file is fetched — so this answers
+         * the MIME half only; `resolveUsableProviderBinding` re-asks with the real length and falls
+         * through if the document is too large, which is the same shape as a PDF exceeding
+         * `maxDocumentBytes`.
+         */
+        return PROVIDER_ONLY_DOCUMENT_MIMES.has(lower) && this.providerDeclaresMimeType(lower);
+    }
+    /**
+     * Whether the configured document-extraction provider declares this MIME type.
+     *
+     * Deliberately narrower than {@link resolveUsableProviderBinding}: no byte length, no logging, and
+     * every failure answers `false`. It is a capability QUESTION asked on a hot path (once per file
+     * per sweep), not the resolution that precedes an actual extraction.
+     */
+    providerDeclaresMimeType(mimeType) {
+        if (!this.documentExtraction)
+            return false;
+        try {
+            const resolution = this.documentExtraction.resolveBinding(DocumentExtraction_Applicability.DOCUMENT_TEXT_DEFAULT);
+            if (resolution.configured === false)
+                return false;
+            return resolution.binding.supportedMimeTypes.includes(mimeType);
+        }
+        catch {
+            // A misconfigured provider is not a supported type. The loud version of this is logged where
+            // extraction is actually attempted; answering `false` here keeps the corpus skip quiet.
+            return false;
+        }
     }
     async extractFromText(stream, fileId) {
         // Plain text keeps its ZERO-dependency path deliberately: routing text/* through a PDF parser
@@ -301,15 +382,41 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
         const binding = options?.allowProviderExtraction === true
             ? this.resolveUsableProviderBinding(fileId, mimeType, buffer.length)
             : undefined;
+        /*
+         * LEGACY BINARY Office only (#434). These are OLE2 compound files, so neither local library
+         * below can serve them — handing a `.doc` to pdf.js or to the ZIP+XML extractor produces an
+         * unreadable-bytes classification that blames the document for a capability the deployment does
+         * not have.
+         *
+         * The six ZIP+XML types deliberately do NOT reach this branch any more: they have a local
+         * library, so a missing binding degrades them to it rather than refusing.
+         *
+         * `isExtractionSupported` already asked the provider whether it declares this MIME, so reaching
+         * here without a binding means the answer changed between the two questions (the byte length
+         * exceeded `maxDocumentBytes`, or the provider became unresolvable). Refusing names that
+         * precisely instead of mislabelling the file.
+         */
+        if (!binding && PROVIDER_ONLY_DOCUMENT_MIMES.has(mimeType.toLowerCase())) {
+            throw this.errorBuilder.buildError(ErrorType.VALIDATION, undefined, {
+                customMessageReference: ErrorCustomMessageReference.FILE_TEXT_EXTRACTION_UNSUPPORTED_TYPE,
+                context: {
+                    // diagnostic (log-only)
+                    fileId,
+                    mimeType,
+                    byteLength: buffer.length,
+                    message: 'This document type is extractable only through a configured document-extraction provider, and none can serve it',
+                },
+            });
+        }
         if (!binding && options?.allowProviderExtraction !== true) {
             // Rung 0 — the field did not opt in. Not a fallback and not a degradation: this is the
             // declared behaviour of a field whose author left `ocr` off (or wrote `ocr: false`), and it
             // is where the overwhelming majority of documents belong, because a born-digital PDF gains
             // nothing from OCR and would cost per page.
-            return await this.extractFromPDF(buffer, fileId, 'local-library (field did not opt into provider extraction)');
+            return await this.extractLocally(buffer, fileId, mimeType, 'local-library (field did not opt into provider extraction)');
         }
         if (!binding) {
-            return await this.extractFromPDF(buffer, fileId, 'local-library (no usable provider)');
+            return await this.extractLocally(buffer, fileId, mimeType, 'local-library (no usable provider)');
         }
         try {
             const extracted = await this.documentExtraction.extractDocumentText({
@@ -348,7 +455,7 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
                 providerRef: binding.providerRef,
                 pageCount: extracted.pageCount,
             });
-            return await this.extractFromPDF(buffer, fileId, 'local-library (provider returned no text)');
+            return await this.extractLocally(buffer, fileId, mimeType, 'local-library (provider returned no text)');
         }
         catch (providerError) {
             if (providerError instanceof DocumentExtractionError
@@ -356,7 +463,7 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
                 // Rung 6 — the corpus-protecting branch.
                 let localResult;
                 try {
-                    localResult = await this.extractFromPDF(buffer, fileId, 'local-library (provider unavailable)');
+                    localResult = await this.extractLocally(buffer, fileId, mimeType, 'local-library (provider unavailable)');
                 }
                 catch {
                     // The local library ALSO failed. Deliberately swallowed: the provider fault is the more
@@ -389,7 +496,7 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
                 providerRef: binding.providerRef,
                 reason: providerError instanceof Error ? providerError.message : String(providerError),
             });
-            return await this.extractFromPDF(buffer, fileId, 'local-library (provider could not read the document)');
+            return await this.extractLocally(buffer, fileId, mimeType, 'local-library (provider could not read the document)');
         }
     }
     /**
@@ -464,6 +571,58 @@ let FileTextExtractionBackendService = class FileTextExtractionBackendService {
      * which rung of that seam routed here and is logged, so an operator reading the logs can tell
      * "no provider is configured" from "the provider was down" without inspecting configuration.
      */
+    /**
+     * The LOCAL-LIBRARY rung of {@link extractFromDocument}'s ladder, dispatched by MIME.
+     *
+     * Two libraries sit at this rung and they are not interchangeable: `pdfjs-dist` reads a PDF's
+     * text layer, and the ZIP+XML extractor reads OOXML / OpenDocument parts. Handing a document to
+     * the wrong one produces an unreadable-bytes classification that blames the FILE for a routing
+     * mistake — which is exactly the failure the provider-only era's refusal existed to avoid, so it
+     * must not reappear now that there are two locals instead of one.
+     *
+     * One dispatcher rather than a branch at each of the five call sites, because the five rungs are
+     * the same decision reached for five different reasons; branching at each would be five places to
+     * forget the office arm the day a sixth rung appears.
+     */
+    async extractLocally(buffer, fileId, mimeType, reason) {
+        if (LOCALLY_EXTRACTABLE_DOCUMENT_MIMES.has(mimeType.toLowerCase())) {
+            return await this.extractFromOfficeDocument(buffer, fileId, mimeType.toLowerCase(), reason);
+        }
+        return await this.extractFromPDF(buffer, fileId, reason);
+    }
+    /**
+     * OOXML / OpenDocument text, read locally (#434).
+     *
+     * `partCount` is reported as `pageCount` deliberately: for a `.pptx` it IS the slide count, and
+     * for the others it is the closest honest analogue of "how many units of this document were
+     * read". Inventing a page count by dividing characters would be a number an operator could
+     * reasonably act on and that means nothing.
+     *
+     * An {@link OfficeDocumentExtractionError} is rethrown UNTOUCHED so the caller's classifier sees
+     * it: every one of its failures is a statement about the document's own bytes — not a readable
+     * ZIP, larger than the decompression bound, or holding none of the parts its type keeps text in
+     * — which is precisely the unreadable-document bucket, and not the provider-unavailable one that
+     * must protect an existing corpus.
+     */
+    async extractFromOfficeDocument(buffer, fileId, mimeType, reason) {
+        const extracted = await extractOfficeDocumentText(buffer, mimeType);
+        this.logDebug('Extracted text from an Office/OpenDocument file with the local library', {
+            fileId,
+            mimeType,
+            reason,
+            partCount: extracted.partCount,
+            characterCount: extracted.text.length,
+        });
+        return {
+            text: extracted.text,
+            metadata: {
+                pageCount: extracted.partCount,
+                characterCount: extracted.text.length,
+                wordCount: extracted.text.split(/\s+/).filter(Boolean).length,
+                extractionPath: 'local-library',
+            },
+        };
+    }
     async extractFromPDF(buffer, fileId, reason) {
         const pdfjs = await import('pdfjs-dist/legacy/build/pdf.mjs');
         const loadingTask = pdfjs.getDocument({

package/dist/esm/files/services/file-text-extraction.backend.service.js.map

@@ -1 +1 @@
-{"version":3,"file":"file-text-extraction.backend.service.js","sourceRoot":"","sources":["../../../../../src/files/services/file-text-extraction.backend.service.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2DG;;;;;;;;;;;;;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AAEzD,OAAO,EAAE,aAAa,EAAE,MAAM,aAAa,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,SAAS,EAAE,2BAA2B,EAAa,MAAM,uBAAuB,CAAC;AAC1F,OAAO,EAAE,kBAAkB,EAAE,MAAM,gDAAgD,CAAC;AAEpF,OAAO,EAAE,0BAA0B,EAAE,MAAM,kEAAkE,CAAC;AAE9G,OAAO,EAAE,+BAA+B,EAAE,MAAM,sEAAsE,CAAC;AACvH,OAAO,EAAE,mBAAmB,EAAE,MAAM,yDAAyD,CAAC;AAC9F,OAAO,EACL,gCAAgC,EAChC,uBAAuB,EACvB,8BAA8B,GAE/B,MAAM,oFAAoF,CAAC;AAC5F,OAAO,EAAE,gCAAgC,EAAE,MAAM,sCAAsC,CAAC;AAuCxF,MAAM,kBAAkB,GAAG,CAAC,OAAO,CAAC,CAAC;AACrC,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC;IAC9B,kBAAkB;IAClB,iBAAiB;IACjB,uBAAuB;IACvB,wBAAwB;IACxB,wBAAwB;IACxB,oBAAoB;IACpB,kBAAkB;IAClB,kBAAkB;IAClB,qBAAqB;IACrB,kBAAkB;CACnB,CAAC,CAAC;AACH,MAAM,QAAQ,GAAG,iBAAiB,CAAC;AAEnC,MAAM,aAAa,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;AAErD;;;;;;;;;;;;;;;;;;;GAmBG;AACH,IAAI,0BAA8C,CAAC;AACnD,SAAS,2BAA2B;IAClC,IAAI,0BAA0B,KAAK,SAAS,EAAE,CAAC;QAC7C,8FAA8F;QAC9F,0BAA0B,GAAG,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,OAAO,CAAC,yBAAyB,CAAC,CAAC,EAAE,gBAAgB,CAAC,GAAG,GAAG,CAAC;IACvH,CAAC;IACD,OAAO,0BAA0B,CAAC;AACpC,CAAC;AAGM,IAAM,gCAAgC,GAAtC,MAAM,gCAAgC;IAKxB;IAEA;IAEA;IAGA;IASA;IApBT,QAAQ,CAA+D;IAEjF,YAEmB,MAAqB,EAErB,YAAwC,EAExC,iBAAkD,EAGlD,YAAiC,EASjC,kBAAqD;QAhBrD,WAAM,GAAN,MAAM,CAAe;QAErB,iBAAY,GAAZ,YAAY,CAA4B;QAExC,sBAAiB,GAAjB,iBAAiB,CAAiC;QAGlD,iBAAY,GAAZ,YAAY,CAAqB;QASjC,uBAAkB,GAAlB,kBAAkB,CAAmC;QAEtE,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,kCAAkC,CAAC,CAAC;QAC5E,IAAI,CAAC,MAAM,CAAC,mBAAmB,CAAC,kCAAkC,EAAE,KAAK,CAAC,CAAC;IAC7E,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CACf,IAAU,EACV,gBAAuC,EACvC,OAAmC;QAEnC,MAAM,QAAQ,GAAG,IAAI,CAAC,QAAQ,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC;QAEpD,IAAI,CAAC,IAAI,CAAC,qBAAqB,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC1C,MAAM,IAAI,CAAC,YAAY,CAAC,UAAU,CAChC,SAAS,CAAC,UAAU,EACpB,gBAAgB,EAChB;gBACE,sBAAsB,EAAE,2BAA2B,CAAC,qCAAqC;gBACzF,OAAO,EAAE;oBACP,wBAAwB;oBACxB,MAAM,EAAE,IAAI,CAAC,MAAM;oBACnB,QAAQ;iBACT;aACF,CACF,CAAC;QACJ,CAAC;QAED,wEAAwE;QACxE,0EAA0E;QAC1E,2EAA2E;QAC3E,4BAA4B;QAC5B,EAAE;QACF,gFAAgF;QAChF,gFAAgF;QAChF,8EAA8E;QAC9E,oFAAoF;QACpF,MAAM,QAAQ,GAAG,IAAI,CAAC,YAAY,CAAC,yBAAyB,CAAC,IAAI,EAAE,gBAAgB,CAAC,CAAC;QAErF,MAAM,cAAc,GAAG,MAAM,QAAQ,CAAC,QAAQ,CAC5C,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,eAAe,EAAE,IAAI,CAAC,eAAe,EAAE,UAAU,EAAE,IAAI,CAAC,UAAU,EAAE,EAC3F,gBAAgB,CACjB,CAAC;QAEF,IAAI,CAAC;YACH,IAAI,CAAC;gBACH,IAAI,QAAQ,KAAK,QAAQ,EAAE,CAAC;oBAC1B,OAAO,MAAM,IAAI,CAAC,mBAAmB,CAAC,cAAc,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,OAAO,CAAC,CAAC;gBAC/F,CAAC;gBACD,OAAO,MAAM,IAAI,CAAC,eAAe,CAAC,cAAc,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;YACxE,CAAC;YAAC,OAAO,UAAU,EAAE,CAAC;gBACpB,oFAAoF;gBACpF,yFAAyF;gBACzF,wFAAwF;gBACxF,yFAAyF;gBACzF,wFAAwF;gBACxF,gCAAgC;gBAChC,IACE,UAAU,YAAY,uBAAuB;uBAC1C,UAAU,CAAC,YAAY,KAAK,8BAA8B,CAAC,oBAAoB,EAClF,CAAC;oBACD,MAAM,UAAU,CAAC;gBACnB,CAAC;gBACD,uFAAuF;gBACvF,sDAAsD;gBACtD,EAAE;gBACF,2FAA2F;gBAC3F,uFAAuF;gBACvF,4FAA4F;gBAC5F,wFAAwF;gBACxF,sCAAsC;gBACtC,EAAE;gBACF,4FAA4F;gBAC5F,4FAA4F;gBAC5F,qDAAqD;gBACrD,EAAE;gBACF,4FAA4F;gBAC5F,yFAAyF;gBACzF,yBAAyB;gBACzB,MAAM,IAAI,CAAC,YAAY,CAAC,UAAU,CAAC,SAAS,CAAC,UAAU,EAAE,gBAAgB,EAAE;oBACzE,sBAAsB,EAAE,2BAA2B,CAAC,2BAA2B;oBAC/E,OAAO,EAAE;wBACP,wBAAwB;wBACxB,MAAM,EAAE,IAAI,CAAC,MAAM;wBACnB,QAAQ;wBACR,KAAK,EAAE,UAAU,YAAY,KAAK,CAAC,CAAC,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC;qBAC7E;iBACF,CAAC,CAAC;YACL,CAAC;QACH,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC;gBACrC,cAAc,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;YAClC,CAAC;QACH,CAAC;IACH,CAAC;IAED;;OAEG;IACH,qBAAqB,CAAC,QAAgB;QACpC,MAAM,KAAK,GAAG,QAAQ,CAAC,WAAW,EAAE,CAAC;QACrC,IAAI,KAAK,KAAK,QAAQ;YAAE,OAAO,IAAI,CAAC;QACpC,IAAI,eAAe,CAAC,GAAG,CAAC,KAAK,CAAC;YAAE,OAAO,IAAI,CAAC;QAC5C,OAAO,kBAAkB,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,KAAK,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC;IACvE,CAAC;IAEO,KAAK,CAAC,eAAe,CAAC,MAAgB,EAAE,MAAc;QAC5D,8FAA8F;QAC9F,2FAA2F;QAC3F,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;QAElE,IAAI,CAAC,QAAQ,CAAC,qCAAqC,EAAE;YACnD,MAAM;YACN,cAAc,EAAE,IAAI,CAAC,MAAM;SAC5B,CAAC,CAAC;QAEH,OAAO;YACL,IAAI;YACJ,QAAQ,EAAE;gBACR,cAAc,EAAE,IAAI,CAAC,MAAM;gBAC3B,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;aACpD;SACF,CAAC;IACJ,CAAC;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OAoDG;IACK,KAAK,CAAC,mBAAmB,CAC/B,MAAgB,EAChB,MAAc,EACd,QAAgB,EAChB,OAAmC;QAEnC,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;QAChD,MAAM,OAAO,GAAG,OAAO,EAAE,uBAAuB,KAAK,IAAI;YACvD,CAAC,CAAC,IAAI,CAAC,4BAA4B,CAAC,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,MAAM,CAAC;YACpE,CAAC,CAAC,SAAS,CAAC;QAEd,IAAI,CAAC,OAAO,IAAI,OAAO,EAAE,uBAAuB,KAAK,IAAI,EAAE,CAAC;YAC1D,uFAAuF;YACvF,4FAA4F;YAC5F,2FAA2F;YAC3F,4CAA4C;YAC5C,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,4DAA4D,CAAC,CAAC;QACjH,CAAC;QAED,IAAI,CAAC,OAAO,EAAE,CAAC;YACb,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,oCAAoC,CAAC,CAAC;QACzF,CAAC;QAED,IAAI,CAAC;YACH,MAAM,SAAS,GAAG,MAAM,IAAI,CAAC,kBAAmB,CAAC,mBAAmB,CAAC;gBACnE,QAAQ,EAAE,MAAM;gBAChB,QAAQ;gBACR,aAAa,EAAE,gCAAgC,CAAC,qBAAqB;gBACrE,MAAM;aACP,CAAC,CAAC;YAEH,IAAI,SAAS,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACrC,IAAI,CAAC,QAAQ,CAAC,gEAAgE,EAAE;oBAC9E,MAAM;oBACN,WAAW,EAAE,OAAO,CAAC,WAAW;oBAChC,KAAK,EAAE,SAAS,CAAC,KAAK;oBACtB,SAAS,EAAE,SAAS,CAAC,SAAS;oBAC9B,cAAc,EAAE,SAAS,CAAC,IAAI,CAAC,MAAM;iBACtC,CAAC,CAAC;gBACH,OAAO;oBACL,IAAI,EAAE,SAAS,CAAC,IAAI;oBACpB,QAAQ,EAAE;wBACR,SAAS,EAAE,SAAS,CAAC,SAAS;wBAC9B,cAAc,EAAE,SAAS,CAAC,IAAI,CAAC,MAAM;wBACrC,SAAS,EAAE,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;wBAC7D,cAAc,EAAE,UAAU;wBAC1B,qBAAqB,EAAE,OAAO,CAAC,WAAW;wBAC1C,eAAe,EAAE,SAAS,CAAC,KAAK;wBAChC,GAAG,CAAC,SAAS,CAAC,KAAK,CAAC,cAAc,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,sBAAsB,EAAE,SAAS,CAAC,KAAK,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;qBACpH;iBACF,CAAC;YACJ,CAAC;YAED,6FAA6F;YAC7F,8FAA8F;YAC9F,2FAA2F;YAC3F,+CAA+C;YAC/C,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,mFAAmF,EAAE;gBACpG,MAAM;gBACN,WAAW,EAAE,OAAO,CAAC,WAAW;gBAChC,SAAS,EAAE,SAAS,CAAC,SAAS;aAC/B,CAAC,CAAC;YACH,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,2CAA2C,CAAC,CAAC;QAChG,CAAC;QAAC,OAAO,aAAa,EAAE,CAAC;YACvB,IACE,aAAa,YAAY,uBAAuB;mBAC7C,aAAa,CAAC,YAAY,KAAK,8BAA8B,CAAC,oBAAoB,EACrF,CAAC;gBACD,yCAAyC;gBACzC,IAAI,WAAiD,CAAC;gBACtD,IAAI,CAAC;oBACH,WAAW,GAAG,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,sCAAsC,CAAC,CAAC;gBAClG,CAAC;gBAAC,MAAM,CAAC;oBACP,wFAAwF;oBACxF,qFAAqF;oBACrF,wFAAwF;oBACxF,qDAAqD;oBACrD,WAAW,GAAG,SAAS,CAAC;gBAC1B,CAAC;gBAED,IAAI,WAAW,IAAI,WAAW,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBACtD,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,4FAA4F,EAAE;wBAC7G,MAAM;wBACN,WAAW,EAAE,OAAO,CAAC,WAAW;wBAChC,UAAU,EAAE,aAAa,CAAC,UAAU;wBACpC,MAAM,EAAE,aAAa,CAAC,OAAO;wBAC7B,cAAc,EAAE,WAAW,CAAC,IAAI,CAAC,MAAM;qBACxC,CAAC,CAAC;oBACH,OAAO,WAAW,CAAC;gBACrB,CAAC;gBAED,IAAI,CAAC,MAAM,CAAC,IAAI,CACd,6IAA6I,EAC7I,EAAE,MAAM,EAAE,WAAW,EAAE,OAAO,CAAC,WAAW,EAAE,UAAU,EAAE,aAAa,CAAC,UAAU,EAAE,MAAM,EAAE,aAAa,CAAC,OAAO,EAAE,CAClH,CAAC;gBACF,wFAAwF;gBACxF,4FAA4F;gBAC5F,MAAM,aAAa,CAAC;YACtB,CAAC;YAED,0FAA0F;YAC1F,2FAA2F;YAC3F,sFAAsF;YACtF,yFAAyF;YACzF,IAAI,CAAC,QAAQ,CAAC,qFAAqF,EAAE;gBACnG,MAAM;gBACN,WAAW,EAAE,OAAO,CAAC,WAAW;gBAChC,MAAM,EAAE,aAAa,YAAY,KAAK,CAAC,CAAC,CAAC,aAAa,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC;aACvF,CAAC,CAAC;YACH,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,sDAAsD,CAAC,CAAC;QAC3G,CAAC;IACH,CAAC;IAED;;;;;;;;;OASG;IACK,4BAA4B,CAClC,MAAc,EACd,QAAgB,EAChB,UAAkB;QAElB,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,CAAC;YAC7B,OAAO,SAAS,CAAC;QACnB,CAAC;QACD,IAAI,CAAC;YACH,MAAM,UAAU,GAAG,IAAI,CAAC,kBAAkB,CAAC,cAAc,CAAC,gCAAgC,CAAC,qBAAqB,CAAC,CAAC;YAClH,6FAA6F;YAC7F,0FAA0F;YAC1F,yFAAyF;YACzF,IAAI,UAAU,CAAC,UAAU,KAAK,KAAK,EAAE,CAAC;gBACpC,IAAI,CAAC,QAAQ,CAAC,sEAAsE,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,CAAC,MAAM,EAAE,CAAC,CAAC;gBAC7H,OAAO,SAAS,CAAC;YACnB,CAAC;YACD,IAAI,CAAC,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,UAAU,CAAC,OAAO,EAAE,QAAQ,EAAE,UAAU,CAAC,EAAE,CAAC;gBAChF,IAAI,CAAC,QAAQ,CAAC,8FAA8F,EAAE;oBAC5G,MAAM;oBACN,QAAQ;oBACR,UAAU;oBACV,gBAAgB,EAAE,UAAU,CAAC,OAAO,CAAC,gBAAgB;oBACrD,WAAW,EAAE,UAAU,CAAC,OAAO,CAAC,WAAW;iBAC5C,CAAC,CAAC;gBACH,OAAO,SAAS,CAAC;YACnB,CAAC;YACD,OAAO,UAAU,CAAC,OAAO,CAAC;QAC5B,CAAC;QAAC,OAAO,eAAe,EAAE,CAAC;YACzB,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,qGAAqG,EAAE;gBACtH,MAAM;gBACN,MAAM,EAAE,eAAe,YAAY,KAAK,CAAC,CAAC,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,eAAe,CAAC;aAC7F,CAAC,CAAC;YACH,OAAO,SAAS,CAAC;QACnB,CAAC;IACH,CAAC;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;OA4BG;IACK,KAAK,CAAC,cAAc,CAAC,MAAc,EAAE,MAAc,EAAE,MAAc;QACzE,MAAM,KAAK,GAAG,MAAM,MAAM,CAAC,iCAAiC,CAAC,CAAC;QAE9D,MAAM,WAAW,GAAG,KAAK,CAAC,WAAW,CAAC;YACpC,kFAAkF;YAClF,yFAAyF;YACzF,sFAAsF;YACtF,0FAA0F;YAC1F,yFAAyF;YACzF,eAAe;YACf,IAAI,EAAE,IAAI,UAAU,CAAC,MAAM,CAAC;YAC5B,uFAAuF;YACvF,sEAAsE;YACtE,EAAE;YACF,4FAA4F;YAC5F,qFAAqF;YACrF,8EAA8E;YAC9E,EAAE;YACF,sFAAsF;YACtF,yFAAyF;YACzF,0EAA0E;YAC1E,EAAE;YACF,2FAA2F;YAC3F,sFAAsF;YACtF,4FAA4F;YAC5F,0CAA0C;YAC1C,cAAc,EAAE,KAAK;YACrB,mBAAmB,EAAE,2BAA2B,EAAE;SACnD,CAAC,CAAC;QAEH,IAAI,CAAC;YACH,MAAM,WAAW,GAAG,MAAM,WAAW,CAAC,OAAO,CAAC;YAC9C,MAAM,SAAS,GAAa,EAAE,CAAC;YAE/B,KAAK,IAAI,UAAU,GAAG,CAAC,EAAE,UAAU,IAAI,WAAW,CAAC,QAAQ,EAAE,UAAU,EAAE,EAAE,CAAC;gBAC1E,MAAM,IAAI,GAAG,MAAM,WAAW,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;gBACnD,IAAI,CAAC;oBACH,MAAM,WAAW,GAAG,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;oBAChD,qFAAqF;oBACrF,EAAE;oBACF,kFAAkF;oBAClF,yFAAyF;oBACzF,wFAAwF;oBACxF,sFAAsF;oBACtF,gEAAgE;oBAChE,EAAE;oBACF,uFAAuF;oBACvF,uFAAuF;oBACvF,8DAA8D;oBAC9D,IAAI,QAAQ,GAAG,EAAE,CAAC;oBAClB,KAAK,MAAM,IAAI,IAAI,WAAW,CAAC,KAAK,EAAE,CAAC;wBACrC,IAAI,CAAC,CAAC,KAAK,IAAI,IAAI,CAAC;4BAAE,SAAS;wBAC/B,QAAQ,IAAI,IAAI,CAAC,GAAG,CAAC;wBACrB,IAAI,IAAI,CAAC,MAAM;4BAAE,QAAQ,IAAI,IAAI,CAAC;oBACpC,CAAC;oBACD,SAAS,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;gBAC3B,CAAC;wBAAS,CAAC;oBACT,IAAI,CAAC,OAAO,EAAE,CAAC;gBACjB,CAAC;YACH,CAAC;YAED,MAAM,IAAI,GAAG,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAElC,IAAI,CAAC,QAAQ,CAAC,uDAAuD,EAAE;gBACrE,MAAM;gBACN,SAAS,EAAE,WAAW,CAAC,QAAQ;gBAC/B,cAAc,EAAE,IAAI,CAAC,MAAM;gBAC3B,MAAM;aACP,CAAC,CAAC;YAEH,OAAO;gBACL,IAAI;gBACJ,QAAQ,EAAE;oBACR,SAAS,EAAE,WAAW,CAAC,QAAQ;oBAC/B,cAAc,EAAE,IAAI,CAAC,MAAM;oBAC3B,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;oBACnD,wFAAwF;oBACxF,oFAAoF;oBACpF,cAAc,EAAE,eAAe;oBAC/B,oBAAoB,EAAE,MAAM;iBAC7B;aACF,CAAC;QACJ,CAAC;gBAAS,CAAC;YACT,8FAA8F;YAC9F,oFAAoF;YACpF,EAAE;YACF,8FAA8F;YAC9F,wFAAwF;YACxF,6FAA6F;YAC7F,6DAA6D;YAC7D,MAAM,WAAW,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;QACrD,CAAC;IACH,CAAC;IAED,qFAAqF;IAC7E,KAAK,CAAC,aAAa,CAAC,MAAgB;QAC1C,MAAM,MAAM,GAAa,EAAE,CAAC;QAC5B,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;YACjC,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;QACnE,CAAC;QACD,OAAO,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IAC/B,CAAC;CACF,CAAA;AA7eY,gCAAgC;IAD5C,UAAU,EAAE;IAKR,WAAA,MAAM,CAAC,kBAAkB,CAAC,MAAM,CAAC,CAAA;IAEjC,WAAA,MAAM,CAAC,kBAAkB,CAAC,mBAAmB,CAAC,CAAA;IAE9C,WAAA,MAAM,CAAC,kBAAkB,CAAC,iBAAiB,CAAC,CAAA;IAG5C,WAAA,MAAM,CAAC,kBAAkB,CAAC,mBAAmB,CAAC,CAAA;IAQ9C,WAAA,MAAM,CAAC,kBAAkB,CAAC,yBAAyB,CAAC,CAAA;IACpD,WAAA,QAAQ,EAAE,CAAA;+CAboB,0BAA0B;QAErB,+BAA+B;QAGpC,mBAAmB;QASZ,gCAAgC;GArB7D,gCAAgC,CA6e5C","sourcesContent":["/**\n * File Text Extraction Backend Service\n *\n * Extracts text content from files. Retained through the Weaviate-lane demolition (2026-08-26):\n * this is the file-side TEXT PROVIDER the resource-native RAG capability reuses in Phase 1 of\n * `.claude/plans/resource-rag-capability.md` (extraction knobs like extractText/ocr stay file-side).\n *\n * Supported formats (first slice):\n * - Plain text (text/plain, text/markdown, text/csv, application/json, etc.)\n * - PDF (via `pdfjs-dist` — dynamically imported, but an INSTALLED dependency)\n *\n * ## Why `pdfjs-dist` and not `pdf-parse` (decided 2026-08-27)\n *\n * The header used to advertise PDF support through `pdf-parse`, which was in NO manifest and NO\n * lockfile anywhere in the repository. So every PDF that reached a `.ragSource()` field threw\n * `'pdf-parse is required for PDF text extraction'` — and because the RAG sweep treats a throw as a\n * failed ingestion and re-visits hourly, that was a permanent retry loop, not a one-off failure.\n * The advertised capability had never worked. Both halves of that are now fixed: the library is a\n * real dependency, and the sweep no longer strands on an unextractable file.\n *\n * `pdf-parse` was reconsidered rather than assumed obsolete — its modern release (2.4.5, Oct 2025)\n * is maintained, contrary to the reputation of the long-abandoned 1.1.x line. It was still rejected,\n * on dependency shape rather than hygiene: it requires `@napi-rs/canvas` (a NATIVE addon, per-platform\n * binaries, and this repository's own toolchain baseline records what a stale C++ toolchain costs)\n * and it vendors `pdfjs-dist@5.4.296` — a SECOND copy at a different major from the `^6.2.108` this\n * repository already pins centrally. `pdfjs-dist` used directly has zero required dependencies\n * (`@napi-rs/canvas` is OPTIONAL there and is needed only for RENDERING, which text extraction never\n * does), is Mozilla-maintained, and matches the `^6.2.108` the root `pnpm.overrides` already pins\n * for the PDF viewer preset. One library, one major, no native build.\n *\n * (The `pdf-parse` dependency facts above were read from the npm registry on 2026-08-27 — `npm view\n * pdf-parse dependencies` reports `pdfjs-dist@5.4.296` and `@napi-rs/canvas@0.1.80`. They are not\n * checkable from inside this repository, so the provenance is recorded rather than assumed.)\n *\n * ## The version constraint is a RANGE, not the usual `*`\n *\n * Most dependencies here are declared `*` because the root `pnpm.overrides` is the single version\n * authority. That works inside the monorepo and NOT for a generated application, which inherits no\n * root override — and `pdfjs-dist` is not in `propagated-engine-overrides.json`. Under `*` an\n * application would happily satisfy this import with the `4.10.38` its PDF viewer already pulls in\n * transitively, against which this file's API assumptions are false (`loadingTask.destroy`, the\n * absent `exports` map that makes both the deep `legacy/build` specifier and the `package.json`\n * resolution legal). `^6.2.108` states what the CODE requires. Precedent: `@react-pdf/renderer`\n * carries a real range in this same manifest for the same reason.\n *\n * ## Provider-backed extraction (OCR) — the selection seam\n *\n * A PDF is served by the application's `EngineCapability.AI_DOCUMENT_EXTRACTION` provider when one\n * is configured and can handle the bytes, and by the local library otherwise. `extractFromDocument`\n * owns that decision and documents the full precedence, including the branch that protects an\n * existing RAG corpus from a provider outage. Plain text never touches a provider.\n *\n * Future:\n * - Office documents (docx via mammoth, xlsx/pptx via dedicated libs)\n * - Audio/video transcript extraction\n *\n * The service operates against the storage provider contract — it downloads\n * the file via `FileStorageRouterBackendService`, never touching raw S3/local\n * paths directly.\n */\n\nimport { injectable, inject, optional } from 'inversify';\nimport { Readable } from 'stream';\nimport { createRequire } from 'node:module';\nimport { dirname, join } from 'node:path';\nimport { ErrorType, ErrorCustomMessageReference, type File } from '@wildo-ai/saas-models';\nimport { SAAS_SERVICE_TYPES } from '~wildo-backend-lib/inject/models/service-types';\nimport type { LoggerService } from '~wildo-backend-lib/monitoring/logger/services/logger.service';\nimport { ErrorBuilderBackendService } from '~wildo-backend-lib/errors/services/error-builder.backend.service';\nimport { ExecutionContext } from '~wildo-backend-lib/runtime/models/execution-context.backend.schemas';\nimport { FileStorageRouterBackendService } from '~wildo-backend-lib/files/storage/file-storage-router.backend.service';\nimport { FilesBackendService } from '~wildo-backend-lib/files/services/files.backend.service';\nimport {\n  DocumentExtractionBackendService,\n  DocumentExtractionError,\n  DocumentExtractionFailureClass,\n  type DocumentExtractionResolvedBinding,\n} from '~wildo-backend-lib/external-providers/services/document-extraction.backend.service';\nimport { DocumentExtraction_Applicability } from '@wildo-ai/external-connectors-models';\n\n/**\n * Per-call extraction options — today, the caller's answer to \"may this document be sent to an\n * external extraction provider?\".\n *\n * The answer belongs to the CALLER because it is a per-FIELD declaration\n * (`z_file({ textExtraction: { ocr } })`) and this service is handed a `File`, which knows nothing\n * about the field it hangs off. `RagIngestionBackendService` is the layer that holds both, so it\n * is the layer that translates one into the other.\n */\nexport interface FileTextExtractionOptions {\n  /**\n   * Whether provider-backed extraction (OCR) is permitted for this document.\n   *\n   * **DEFAULT FALSE, deliberately, and it is a spending decision rather than a caution.** OCR is\n   * billed per page by an external provider, and `FileTextExtractionOptions.ocr` on the schema is\n   * documented to application developers as off-by-default for exactly that reason. Defaulting\n   * this to true would mean that merely CONFIGURING a provider silently began charging per page\n   * for every document in every corpus — including fields whose author wrote `ocr: false` and\n   * explained why. Wonder Todos' `knowledge-documents` field is precisely that case.\n   *\n   * When false, extraction uses the local text-layer read and nothing leaves the deployment.\n   */\n  allowProviderExtraction?: boolean;\n}\n\nexport interface FileTextExtractionResult {\n  text: string;\n  language?: string;\n  confidence?: number;\n  metadata?: {\n    pageCount?: number;\n    wordCount?: number;\n    characterCount?: number;\n    [key: string]: unknown;\n  };\n}\n\nconst TEXT_MIME_PREFIXES = ['text/'];\nconst TEXT_MIME_EXACT = new Set([\n  'application/json',\n  'application/xml',\n  'application/xhtml+xml',\n  'application/javascript',\n  'application/typescript',\n  'application/x-yaml',\n  'application/yaml',\n  'application/x-sh',\n  'application/graphql',\n  'application/toml',\n]);\nconst PDF_MIME = 'application/pdf';\n\nconst moduleRequire = createRequire(import.meta.url);\n\n/**\n * Absolute path to `pdfjs-dist`'s bundled standard-font data, resolved from the INSTALLED package\n * rather than assumed relative to this file — `saas-backend-lib` runs from `dist/` in every\n * deployed process, so a path relative to the source tree would be wrong everywhere that matters.\n *\n * Without this, pdf.js emits `Warning: UnknownErrorException: Ensure that the 'standardFontDataUrl'\n * API parameter is provided` for any document using the 14 PDF standard fonts (Helvetica, Times,\n * Courier …) — the family ordinary word processors and report generators emit.\n *\n * MEASURED 2026-08-27, because the obvious claim about this is wrong: removing this parameter does\n * NOT break extraction of the committed Helvetica fixture. Text extraction reads the string\n * operands of the content stream and maps them through the font's ENCODING, which pdf.js can build\n * for a standard font without loading its font PROGRAM. So this is supplied to silence a real\n * warning and to keep glyph mapping correct for documents whose encoding genuinely requires the\n * font program — NOT because the common case depends on it. No fixture currently pins a document\n * that would fail without it; do not claim one does.\n *\n * Resolved lazily and memoized: `require.resolve` touches the filesystem, and extraction runs per\n * file inside a sweep that can process many files in one pass.\n */\nlet cachedStandardFontDataPath: string | undefined;\nfunction resolveStandardFontDataPath(): string {\n  if (cachedStandardFontDataPath === undefined) {\n    // The trailing separator is REQUIRED — pdf.js concatenates the font filename onto this value.\n    cachedStandardFontDataPath = join(dirname(moduleRequire.resolve('pdfjs-dist/package.json')), 'standard_fonts') + '/';\n  }\n  return cachedStandardFontDataPath;\n}\n\n@injectable()\nexport class FileTextExtractionBackendService {\n  protected logDebug: (message: string, context?: Record<string, unknown>) => void;\n\n  constructor(\n    @inject(SAAS_SERVICE_TYPES.Logger)\n    private readonly logger: LoggerService,\n    @inject(SAAS_SERVICE_TYPES.ErrorBuilderService)\n    private readonly errorBuilder: ErrorBuilderBackendService,\n    @inject(SAAS_SERVICE_TYPES.FileStorageRouter)\n    private readonly fileStorageRouter: FileStorageRouterBackendService,\n    // Appended LAST deliberately — DI constructor positions are order-sensitive.\n    @inject(SAAS_SERVICE_TYPES.FilesBackendService)\n    private readonly filesService: FilesBackendService,\n    // OPTIONAL, and that is not defensiveness. This service is constructed directly in several\n    // tests and integration harnesses with a fixed positional argument list; a required dependency\n    // appended here would break every one of them, and — worse — a harness that passed `undefined`\n    // to satisfy the arity would look wired while being dead. Optional makes \"no provider\" the\n    // explicit, tested path it genuinely is: an application with no `AI_DOCUMENT_EXTRACTION`\n    // selection behaves identically to one whose container never bound this service.\n    @inject(SAAS_SERVICE_TYPES.DocumentExtractionService)\n    @optional()\n    private readonly documentExtraction?: DocumentExtractionBackendService,\n  ) {\n    this.logDebug = this.logger.debugByType('FileTextExtractionBackendService');\n    this.logger.setDebugTypeLogging('FileTextExtractionBackendService', false);\n  }\n\n  /**\n   * Extract text from a file by downloading it through the storage provider.\n   */\n  async extractText(\n    file: File,\n    executionContext: ExecutionContext<any>,\n    options?: FileTextExtractionOptions,\n  ): Promise<FileTextExtractionResult> {\n    const mimeType = file.mimeType?.toLowerCase() ?? '';\n\n    if (!this.isExtractionSupported(mimeType)) {\n      throw this.errorBuilder.buildError(\n        ErrorType.VALIDATION,\n        executionContext,\n        {\n          customMessageReference: ErrorCustomMessageReference.FILE_TEXT_EXTRACTION_UNSUPPORTED_TYPE,\n          context: {\n            // diagnostic (log-only)\n            fileId: file.fileId,\n            mimeType,\n          },\n        },\n      );\n    }\n\n    // Resolve through FilesBackendService so a LEGACY row that persisted no\n    // `storageProvider` falls back to the default provider instead of failing\n    // extraction. Reimplementing the lookup here is what made this one of five\n    // behaviours for one input.\n    //\n    // Unlike the scan pipeline, the throw is propagated: this runs on a synchronous\n    // request path with no terminal status to strand, so failing loudly is correct.\n    // The error type moves CONFIGURATION -> INTERNAL, matching the reference — an\n    // unregistered persisted id is an engine wiring fault, not an app misconfiguration.\n    const provider = this.filesService.getStorageProviderForFile(file, executionContext);\n\n    const downloadResult = await provider.download(\n      { fileId: file.fileId, storageProvider: file.storageProvider, storageRef: file.storageRef },\n      executionContext,\n    );\n\n    try {\n      try {\n        if (mimeType === PDF_MIME) {\n          return await this.extractFromDocument(downloadResult.stream, file.fileId, mimeType, options);\n        }\n        return await this.extractFromText(downloadResult.stream, file.fileId);\n      } catch (parseError) {\n        // A provider fault that could not be covered by the local library reaches here as a\n        // `DocumentExtractionError` of class PROVIDER_UNAVAILABLE, and it MUST NOT be classified\n        // as unreadable bytes — see `extractFromDocument` for the full reasoning. Rethrowing it\n        // untouched puts it in the same bucket as a storage outage, which is exactly right: both\n        // mean \"we could not read this document RIGHT NOW\", and both must fail the ingestion so\n        // the existing corpus survives.\n        if (\n          parseError instanceof DocumentExtractionError\n          && parseError.failureClass === DocumentExtractionFailureClass.PROVIDER_UNAVAILABLE\n        ) {\n          throw parseError;\n        }\n        // CLASSIFY the failure, because callers must be able to tell two very different things\n        // apart and a raw pdf.js exception does not let them.\n        //\n        // Everything ABOVE this point — resolving the storage provider, downloading the blob — can\n        // fail for INFRASTRUCTURE reasons: an unregistered provider id (a wiring fault), or an\n        // unreachable object store (transient). Those errors propagate untouched, and must, because\n        // a batch consumer that treats them as \"this file is unreadable\" would draw a permanent\n        // conclusion from a temporary outage.\n        //\n        // Everything INSIDE this try is parsing bytes we already hold. A failure here is a property\n        // of the document — corrupt, truncated, password-protected — and re-reading it later cannot\n        // help. Only that class is safe for a sweep to skip.\n        //\n        // The distinction is not cosmetic: `RagIngestionBackendService` deletes a field's chunk set\n        // when a pass yields no text, so mislabelling a storage outage as unreadable bytes would\n        // destroy a live corpus.\n        throw this.errorBuilder.buildError(ErrorType.VALIDATION, executionContext, {\n          customMessageReference: ErrorCustomMessageReference.FILE_TEXT_EXTRACTION_FAILED,\n          context: {\n            // diagnostic (log-only)\n            fileId: file.fileId,\n            mimeType,\n            cause: parseError instanceof Error ? parseError.message : String(parseError),\n          },\n        });\n      }\n    } finally {\n      if (!downloadResult.stream.destroyed) {\n        downloadResult.stream.destroy();\n      }\n    }\n  }\n\n  /**\n   * Check if a MIME type supports text extraction.\n   */\n  isExtractionSupported(mimeType: string): boolean {\n    const lower = mimeType.toLowerCase();\n    if (lower === PDF_MIME) return true;\n    if (TEXT_MIME_EXACT.has(lower)) return true;\n    return TEXT_MIME_PREFIXES.some((prefix) => lower.startsWith(prefix));\n  }\n\n  private async extractFromText(stream: Readable, fileId: string): Promise<FileTextExtractionResult> {\n    // Plain text keeps its ZERO-dependency path deliberately: routing text/* through a PDF parser\n    // or a remote provider would add cost and an egress for bytes that are already the answer.\n    const text = (await this.collectStream(stream)).toString('utf-8');\n\n    this.logDebug('Extracted text from plain text file', {\n      fileId,\n      characterCount: text.length,\n    });\n\n    return {\n      text,\n      metadata: {\n        characterCount: text.length,\n        wordCount: text.split(/\\s+/).filter(Boolean).length,\n      },\n    };\n  }\n\n  /**\n   * THE SELECTION SEAM — provider first when one is configured and usable, local library\n   * otherwise, and a deliberate answer for every way both can fail.\n   *\n   * ## The precedence, and why each rung is where it is\n   *\n   * 0. **The FIELD did not opt into provider extraction** → local library, with no provider\n   *    resolution attempted at all. `z_file({ textExtraction: { ocr } })` is documented to\n   *    application developers as off-by-default because OCR is an external per-page cost, so\n   *    \"an OCR provider is configured somewhere in this application\" must never be sufficient to\n   *    start spending on a field whose author declared `ocr: false`. Two conditions, both\n   *    required — the same composition as `.ragSource()` + `AI_EMBEDDINGS`.\n   * 1. **No provider configured** → local library. Identical to the behaviour before this seam\n   *    existed, and it is the default: an application that ingests only born-digital PDFs never\n   *    needs OCR and never pays for an egress.\n   * 2. **Provider configured but cannot serve THESE bytes** (MIME outside the model's catalog row,\n   *    or larger than the effective ceiling) → local library. Checked with `canServe()`, which is\n   *    a pure catalog lookup: learning this from a rejected request would cost an egress of the\n   *    application's document to discover something already known.\n   * 3. **Provider returns text** → provider wins. This is the whole point — a scanned contract has\n   *    no text layer, so this is the only rung that can produce anything for it.\n   * 4. **Provider returns EMPTY text** → try the local library anyway. Cheap, and the two read\n   *    genuinely different things: an OCR model can return nothing for a page whose text layer\n   *    pdf.js reads perfectly well (a vector-drawn document, say). Preferring \"some text\" over\n   *    \"no text\" needs no further justification.\n   * 5. **Provider says the DOCUMENT is unreadable** (`UNREADABLE_DOCUMENT`) → try the local\n   *    library, then classify as unreadable if it too yields nothing. Permanent by nature, so the\n   *    sweep should skip rather than retry.\n   * 6. **Provider was UNREACHABLE** (`PROVIDER_UNAVAILABLE` — network, 5xx, 429, bad credential,\n   *    timeout) → try the local library. If it yields text, serve that, degraded but correct. If\n   *    it yields NOTHING, **rethrow the provider fault** rather than reporting empty text.\n   *\n   * Rung 6 is the one that matters, and it is the reason this method is not four lines long.\n   *\n   * `RagIngestionBackendService` deletes a field's entire chunk set when an ingestion pass yields\n   * no text. So for a SCANNED document — one that only the provider can read — a provider outage\n   * plus a \"return empty text\" response would silently destroy a live, already-embedded corpus and\n   * report success, and the next sweep after the outage would rebuild it. That is data loss caused\n   * by a transient network fault, which is not an acceptable trade for a tidier code path.\n   * Rethrowing makes the ingestion FAIL, which leaves the existing chunks untouched and lets the\n   * sweep retry next hour — the treatment that already exists for a storage outage, reused because\n   * the two situations are the same situation.\n   *\n   * The distinction is only expressible because the provider service CLASSIFIES its failures; a\n   * seam that caught a bare `Error` could not tell rung 5 from rung 6, and would have to pick one\n   * wrong answer for the other.\n   *\n   * ## What this method never does\n   *\n   * It never routes `text/*` anywhere near a provider — plain text keeps its zero-dependency path,\n   * because the bytes already ARE the answer and sending them would be a pointless egress of\n   * application content. That gate lives in `extractText` above, not here.\n   */\n  private async extractFromDocument(\n    stream: Readable,\n    fileId: string,\n    mimeType: string,\n    options?: FileTextExtractionOptions,\n  ): Promise<FileTextExtractionResult> {\n    const buffer = await this.collectStream(stream);\n    const binding = options?.allowProviderExtraction === true\n      ? this.resolveUsableProviderBinding(fileId, mimeType, buffer.length)\n      : undefined;\n\n    if (!binding && options?.allowProviderExtraction !== true) {\n      // Rung 0 — the field did not opt in. Not a fallback and not a degradation: this is the\n      // declared behaviour of a field whose author left `ocr` off (or wrote `ocr: false`), and it\n      // is where the overwhelming majority of documents belong, because a born-digital PDF gains\n      // nothing from OCR and would cost per page.\n      return await this.extractFromPDF(buffer, fileId, 'local-library (field did not opt into provider extraction)');\n    }\n\n    if (!binding) {\n      return await this.extractFromPDF(buffer, fileId, 'local-library (no usable provider)');\n    }\n\n    try {\n      const extracted = await this.documentExtraction!.extractDocumentText({\n        document: buffer,\n        mimeType,\n        applicability: DocumentExtraction_Applicability.DOCUMENT_TEXT_DEFAULT,\n        fileId,\n      });\n\n      if (extracted.text.trim().length > 0) {\n        this.logDebug('Extracted text via the configured document-extraction provider', {\n          fileId,\n          providerRef: binding.providerRef,\n          model: extracted.model,\n          pageCount: extracted.pageCount,\n          characterCount: extracted.text.length,\n        });\n        return {\n          text: extracted.text,\n          metadata: {\n            pageCount: extracted.pageCount,\n            characterCount: extracted.text.length,\n            wordCount: extracted.text.split(/\\s+/).filter(Boolean).length,\n            extractionPath: 'provider',\n            extractionProviderRef: binding.providerRef,\n            extractionModel: extracted.model,\n            ...(extracted.usage.pagesProcessed !== undefined ? { providerPagesProcessed: extracted.usage.pagesProcessed } : {}),\n          },\n        };\n      }\n\n      // Rung 4. WARN rather than debug: a provider that read the document and found no text at all\n      // is worth noticing — it is either a genuinely blank document or a model that failed quietly,\n      // and this service's debug channel is OFF by default, which is how the original \"every PDF\n      // throws\" defect stayed invisible for so long.\n      this.logger.warn('Document-extraction provider returned no text — falling back to the local library', {\n        fileId,\n        providerRef: binding.providerRef,\n        pageCount: extracted.pageCount,\n      });\n      return await this.extractFromPDF(buffer, fileId, 'local-library (provider returned no text)');\n    } catch (providerError) {\n      if (\n        providerError instanceof DocumentExtractionError\n        && providerError.failureClass === DocumentExtractionFailureClass.PROVIDER_UNAVAILABLE\n      ) {\n        // Rung 6 — the corpus-protecting branch.\n        let localResult: FileTextExtractionResult | undefined;\n        try {\n          localResult = await this.extractFromPDF(buffer, fileId, 'local-library (provider unavailable)');\n        } catch {\n          // The local library ALSO failed. Deliberately swallowed: the provider fault is the more\n          // informative of the two and is the one whose class the caller branches on, so it is\n          // what gets rethrown below. Losing a pdf.js parse message here costs a log line; losing\n          // the provider's classification would cost a corpus.\n          localResult = undefined;\n        }\n\n        if (localResult && localResult.text.trim().length > 0) {\n          this.logger.warn('Document-extraction provider was unavailable — served this document from the local library', {\n            fileId,\n            providerRef: binding.providerRef,\n            httpStatus: providerError.httpStatus,\n            reason: providerError.message,\n            characterCount: localResult.text.length,\n          });\n          return localResult;\n        }\n\n        this.logger.warn(\n          'Document-extraction provider was unavailable and the local library produced no text — failing the extraction to protect any existing corpus',\n          { fileId, providerRef: binding.providerRef, httpStatus: providerError.httpStatus, reason: providerError.message },\n        );\n        // NOT classified as unreadable bytes. `extractText`'s catch rethrows this untouched, so\n        // the RAG sweep reports FAILED and leaves existing chunks alone — see this method's header.\n        throw providerError;\n      }\n\n      // Rung 5, plus any non-classified throw (a CONFIGURATION error from resolveBinding, say).\n      // Both are treated the same way here on purpose: try the local library, and if it produces\n      // nothing let the failure fall through to `extractText`'s classifier. A configuration\n      // mistake reaching this point has already been logged as itself by the provider service.\n      this.logDebug('Document-extraction provider could not read the document — trying the local library', {\n        fileId,\n        providerRef: binding.providerRef,\n        reason: providerError instanceof Error ? providerError.message : String(providerError),\n      });\n      return await this.extractFromPDF(buffer, fileId, 'local-library (provider could not read the document)');\n    }\n  }\n\n  /**\n   * The usable provider binding for this document, or `undefined` when the local library should\n   * serve — rungs 1 and 2 of the precedence in {@link extractFromDocument}.\n   *\n   * Never throws. A CONFIGURATION fault here (a provider selected but declaring no contract) is\n   * logged and treated as \"no provider\": extraction is an ENHANCEMENT over a local path that\n   * works, so an operator's provider mistake must not take down file extraction for documents the\n   * local library reads perfectly well. The provider service still throws that error loudly on the\n   * `extractDocumentText` path, so the mistake is not silenced — only prevented from cascading.\n   */\n  private resolveUsableProviderBinding(\n    fileId: string,\n    mimeType: string,\n    byteLength: number,\n  ): DocumentExtractionResolvedBinding | undefined {\n    if (!this.documentExtraction) {\n      return undefined;\n    }\n    try {\n      const resolution = this.documentExtraction.resolveBinding(DocumentExtraction_Applicability.DOCUMENT_TEXT_DEFAULT);\n      // `=== false` rather than `!…`: this compiler configuration narrows the literal-discriminant\n      // union on the explicit comparison but not on the truthiness form. Measured here too, not\n      // copied on faith — `!resolution.configured` compiles and then fails to narrow `reason`.\n      if (resolution.configured === false) {\n        this.logDebug('No document-extraction provider configured — using the local library', { fileId, reason: resolution.reason });\n        return undefined;\n      }\n      if (!this.documentExtraction.canServe(resolution.binding, mimeType, byteLength)) {\n        this.logDebug('Configured document-extraction provider cannot serve this document — using the local library', {\n          fileId,\n          mimeType,\n          byteLength,\n          maxDocumentBytes: resolution.binding.maxDocumentBytes,\n          providerRef: resolution.binding.providerRef,\n        });\n        return undefined;\n      }\n      return resolution.binding;\n    } catch (resolutionError) {\n      this.logger.warn('Document-extraction provider is misconfigured — falling back to the local library for this document', {\n        fileId,\n        reason: resolutionError instanceof Error ? resolutionError.message : String(resolutionError),\n      });\n      return undefined;\n    }\n  }\n\n  /**\n   * Extract text from a PDF using the LOCAL `pdfjs-dist` library.\n   *\n   * ## Why the import stays dynamic even though the dependency is installed\n   *\n   * `pdfjs-dist` is a large module graph that most backend processes never touch — only a request\n   * that actually carries a PDF needs it. A dynamic import keeps it out of every other process's\n   * startup cost. What changed is that the import can no longer FAIL for want of a dependency: the\n   * package is declared in this package's manifest, so a resolution failure now means a broken\n   * installation, which is a real fault and is correctly loud.\n   *\n   * The deep specifier targets the `legacy` build deliberately. It is the ESM build pdf.js ships for\n   * non-browser hosts; the default build assumes browser globals. `pdfjs-dist` publishes no\n   * `exports` map, so this deep path is a supported entrypoint rather than a boundary violation.\n   *\n   * ## When a PROVIDER serves this instead\n   *\n   * This local path reads the text layer a PDF already contains, and returns nothing useful for a\n   * SCANNED document (an image of a page carries no text layer). That is what the provider-backed\n   * OCR capability exists to handle, and {@link extractFromDocument} is the seam that chooses\n   * between them. With no `AI_DOCUMENT_EXTRACTION` provider configured, a scanned PDF still\n   * extracts to empty text — which the RAG ingestion path treats as \"contributes no chunks\"\n   * rather than as an error.\n   *\n   * It takes a BUFFER rather than a stream because the selection seam above may need the same\n   * bytes twice (provider first, this second), and a `Readable` is consumed once. `reason` names\n   * which rung of that seam routed here and is logged, so an operator reading the logs can tell\n   * \"no provider is configured\" from \"the provider was down\" without inspecting configuration.\n   */\n  private async extractFromPDF(buffer: Buffer, fileId: string, reason: string): Promise<FileTextExtractionResult> {\n    const pdfjs = await import('pdfjs-dist/legacy/build/pdf.mjs');\n\n    const loadingTask = pdfjs.getDocument({\n      // A COPY, and it is mandatory rather than defensive. pdf.js transfers the backing\n      // ArrayBuffer to its worker port and DETACHES it. For a small allocation Node's `Buffer`\n      // is a window onto a SHARED 8 KB pool, so handing it `buffer` directly would detach a\n      // region other, unrelated buffers are still using. `new Uint8Array(buffer)` copies out of\n      // the pool into a private allocation. Costs 2x peak memory for the document; correctness\n      // is worth it.\n      data: new Uint8Array(buffer),\n      // Not the hardening it looks like — measured against the shipped build 2026-08-27, and\n      // recorded because the plausible reading is wrong in both directions.\n      //\n      // It does NOT stop a document reaching the host's font files: `useSystemFonts` only decides\n      // whether the worker builds a CSS `local()` substitution descriptor, which is string\n      // construction and opens nothing. And in Node it is ALREADY false by default.\n      //\n      // It is set explicitly for the opposite-facing reason: when `useSystemFonts` is true,\n      // `fetchStandardFontData` early-returns null, which would make the `standardFontDataUrl`\n      // below dead. Pinning it false is what keeps that parameter load-bearing.\n      //\n      // (A sibling `isEvalSupported: false` was written here first and removed: pdf.js 6 deleted\n      // that option outright — zero occurrences in its types AND its shipped build — having\n      // removed the eval-based font-compilation path. It type-errors, and would have advertised a\n      // decision the library silently ignored.)\n      useSystemFonts: false,\n      standardFontDataUrl: resolveStandardFontDataPath(),\n    });\n\n    try {\n      const pdfDocument = await loadingTask.promise;\n      const pageTexts: string[] = [];\n\n      for (let pageNumber = 1; pageNumber <= pdfDocument.numPages; pageNumber++) {\n        const page = await pdfDocument.getPage(pageNumber);\n        try {\n          const textContent = await page.getTextContent();\n          // Concatenate items HONOURING `hasEOL`, rather than joining every item with a space.\n          //\n          // pdf.js splits a visual line into several `TextItem`s at kerning and positioning\n          // boundaries, so a blanket `join(' ')` inserts spaces INSIDE words on ordinary documents\n          // and erases line structure entirely. That matters more here than it would in a viewer:\n          // this string is what gets chunked and embedded, so the damage lands in the retrieval\n          // corpus rather than on a screen where a human would notice it.\n          //\n          // `items` also mixes TextItem (which carries `str`) with TextMarkedContent (structural\n          // markers, which do not). Narrowing on the property rather than casting keeps a future\n          // pdf.js item kind from silently stringifying as `undefined`.\n          let pageText = '';\n          for (const item of textContent.items) {\n            if (!('str' in item)) continue;\n            pageText += item.str;\n            if (item.hasEOL) pageText += '\\n';\n          }\n          pageTexts.push(pageText);\n        } finally {\n          page.cleanup();\n        }\n      }\n\n      const text = pageTexts.join('\\n');\n\n      this.logDebug('Extracted text from PDF via the local pdfjs-dist path', {\n        fileId,\n        pageCount: pdfDocument.numPages,\n        characterCount: text.length,\n        reason,\n      });\n\n      return {\n        text,\n        metadata: {\n          pageCount: pdfDocument.numPages,\n          characterCount: text.length,\n          wordCount: text.split(/\\s+/).filter(Boolean).length,\n          // Which path served this extraction, and why. Carried in the RESULT rather than only in\n          // a log line so a caller (and a test) can assert the routing without scraping logs.\n          extractionPath: 'local-library',\n          extractionPathReason: reason,\n        },\n      };\n    } finally {\n      // `destroy()` lives on the LOADING TASK, not on the document proxy — `pdfDocument.destroy` is\n      // not a function in pdfjs-dist 6.x. Skipping this leaks the document's worker port.\n      //\n      // Swallowed deliberately: `destroy()` can itself throw, and a throw from a `finally` REPLACES\n      // the in-flight error. The caller classifies on that original error to decide whether a\n      // document is permanently unreadable or the storage layer is merely down, so losing it would\n      // corrupt exactly the decision this method exists to inform.\n      await loadingTask.destroy().catch(() => undefined);\n    }\n  }\n\n  /** Drain a download stream into a single buffer. Shared by every extraction path. */\n  private async collectStream(stream: Readable): Promise<Buffer> {\n    const chunks: Buffer[] = [];\n    for await (const chunk of stream) {\n      chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk));\n    }\n    return Buffer.concat(chunks);\n  }\n}\n"]}
+{"version":3,"file":"file-text-extraction.backend.service.js","sourceRoot":"","sources":["../../../../../src/files/services/file-text-extraction.backend.service.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiFG;;;;;;;;;;;;;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AAEzD,OAAO,EAAE,aAAa,EAAE,MAAM,aAAa,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,SAAS,EAAE,2BAA2B,EAAa,MAAM,uBAAuB,CAAC;AAC1F,OAAO,EACL,yBAAyB,EACzB,0BAA0B,EAC1B,kCAAkC,GAEnC,MAAM,8EAA8E,CAAC;AACtF,OAAO,EAAE,kBAAkB,EAAE,MAAM,gDAAgD,CAAC;AAEpF,OAAO,EAAE,0BAA0B,EAAE,MAAM,kEAAkE,CAAC;AAE9G,OAAO,EAAE,+BAA+B,EAAE,MAAM,sEAAsE,CAAC;AACvH,OAAO,EAAE,mBAAmB,EAAE,MAAM,yDAAyD,CAAC;AAC9F,OAAO,EACL,gCAAgC,EAChC,uBAAuB,EACvB,8BAA8B,GAE/B,MAAM,oFAAoF,CAAC;AAC5F,OAAO,EAAE,gCAAgC,EAAE,MAAM,sCAAsC,CAAC;AAuCxF,MAAM,kBAAkB,GAAG,CAAC,OAAO,CAAC,CAAC;AACrC,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC;IAC9B,kBAAkB;IAClB,iBAAiB;IACjB,uBAAuB;IACvB,wBAAwB;IACxB,wBAAwB;IACxB,oBAAoB;IACpB,kBAAkB;IAClB,kBAAkB;IAClB,qBAAqB;IACrB,kBAAkB;CACnB,CAAC,CAAC;AACH,MAAM,QAAQ,GAAG,iBAAiB,CAAC;AAEnC;;;;;;;;;;;;;;;GAeG;AACH,MAAM,4BAA4B,GAAwB,0BAA0B,CAAC;AAErF,MAAM,aAAa,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;AAErD;;;;;;;;;;;;;;;;;;;GAmBG;AACH,IAAI,0BAA8C,CAAC;AACnD,SAAS,2BAA2B;IAClC,IAAI,0BAA0B,KAAK,SAAS,EAAE,CAAC;QAC7C,8FAA8F;QAC9F,0BAA0B,GAAG,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,OAAO,CAAC,yBAAyB,CAAC,CAAC,EAAE,gBAAgB,CAAC,GAAG,GAAG,CAAC;IACvH,CAAC;IACD,OAAO,0BAA0B,CAAC;AACpC,CAAC;AAGM,IAAM,gCAAgC,GAAtC,MAAM,gCAAgC;IAKxB;IAEA;IAEA;IAGA;IASA;IApBT,QAAQ,CAA+D;IAEjF,YAEmB,MAAqB,EAErB,YAAwC,EAExC,iBAAkD,EAGlD,YAAiC,EASjC,kBAAqD;QAhBrD,WAAM,GAAN,MAAM,CAAe;QAErB,iBAAY,GAAZ,YAAY,CAA4B;QAExC,sBAAiB,GAAjB,iBAAiB,CAAiC;QAGlD,iBAAY,GAAZ,YAAY,CAAqB;QASjC,uBAAkB,GAAlB,kBAAkB,CAAmC;QAEtE,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,kCAAkC,CAAC,CAAC;QAC5E,IAAI,CAAC,MAAM,CAAC,mBAAmB,CAAC,kCAAkC,EAAE,KAAK,CAAC,CAAC;IAC7E,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CACf,IAAU,EACV,gBAAuC,EACvC,OAAmC;QAEnC,MAAM,QAAQ,GAAG,IAAI,CAAC,QAAQ,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC;QAEpD,IAAI,CAAC,IAAI,CAAC,qBAAqB,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC1C,MAAM,IAAI,CAAC,YAAY,CAAC,UAAU,CAChC,SAAS,CAAC,UAAU,EACpB,gBAAgB,EAChB;gBACE,sBAAsB,EAAE,2BAA2B,CAAC,qCAAqC;gBACzF,OAAO,EAAE;oBACP,wBAAwB;oBACxB,MAAM,EAAE,IAAI,CAAC,MAAM;oBACnB,QAAQ;iBACT;aACF,CACF,CAAC;QACJ,CAAC;QAED,wEAAwE;QACxE,0EAA0E;QAC1E,2EAA2E;QAC3E,4BAA4B;QAC5B,EAAE;QACF,gFAAgF;QAChF,gFAAgF;QAChF,8EAA8E;QAC9E,oFAAoF;QACpF,MAAM,QAAQ,GAAG,IAAI,CAAC,YAAY,CAAC,yBAAyB,CAAC,IAAI,EAAE,gBAAgB,CAAC,CAAC;QAErF,MAAM,cAAc,GAAG,MAAM,QAAQ,CAAC,QAAQ,CAC5C,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,eAAe,EAAE,IAAI,CAAC,eAAe,EAAE,UAAU,EAAE,IAAI,CAAC,UAAU,EAAE,EAC3F,gBAAgB,CACjB,CAAC;QAEF,IAAI,CAAC;YACH,IAAI,CAAC;gBACH,IACE,QAAQ,KAAK,QAAQ;uBAClB,kCAAkC,CAAC,GAAG,CAAC,QAAQ,CAAC;uBAChD,4BAA4B,CAAC,GAAG,CAAC,QAAQ,CAAC,EAC7C,CAAC;oBACD,OAAO,MAAM,IAAI,CAAC,mBAAmB,CAAC,cAAc,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,OAAO,CAAC,CAAC;gBAC/F,CAAC;gBACD,OAAO,MAAM,IAAI,CAAC,eAAe,CAAC,cAAc,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;YACxE,CAAC;YAAC,OAAO,UAAU,EAAE,CAAC;gBACpB,oFAAoF;gBACpF,yFAAyF;gBACzF,wFAAwF;gBACxF,yFAAyF;gBACzF,wFAAwF;gBACxF,gCAAgC;gBAChC,IACE,UAAU,YAAY,uBAAuB;uBAC1C,UAAU,CAAC,YAAY,KAAK,8BAA8B,CAAC,oBAAoB,EAClF,CAAC;oBACD,MAAM,UAAU,CAAC;gBACnB,CAAC;gBACD,uFAAuF;gBACvF,sDAAsD;gBACtD,EAAE;gBACF,2FAA2F;gBAC3F,uFAAuF;gBACvF,4FAA4F;gBAC5F,wFAAwF;gBACxF,sCAAsC;gBACtC,EAAE;gBACF,4FAA4F;gBAC5F,4FAA4F;gBAC5F,qDAAqD;gBACrD,EAAE;gBACF,4FAA4F;gBAC5F,yFAAyF;gBACzF,yBAAyB;gBACzB,MAAM,IAAI,CAAC,YAAY,CAAC,UAAU,CAAC,SAAS,CAAC,UAAU,EAAE,gBAAgB,EAAE;oBACzE,sBAAsB,EAAE,2BAA2B,CAAC,2BAA2B;oBAC/E,OAAO,EAAE;wBACP,wBAAwB;wBACxB,MAAM,EAAE,IAAI,CAAC,MAAM;wBACnB,QAAQ;wBACR,KAAK,EAAE,UAAU,YAAY,KAAK,CAAC,CAAC,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC;qBAC7E;iBACF,CAAC,CAAC;YACL,CAAC;QACH,CAAC;gBAAS,CAAC;YACT,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC;gBACrC,cAAc,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;YAClC,CAAC;QACH,CAAC;IACH,CAAC;IAED;;OAEG;IACH,qBAAqB,CAAC,QAAgB;QACpC,MAAM,KAAK,GAAG,QAAQ,CAAC,WAAW,EAAE,CAAC;QACrC,IAAI,KAAK,KAAK,QAAQ;YAAE,OAAO,IAAI,CAAC;QACpC,IAAI,eAAe,CAAC,GAAG,CAAC,KAAK,CAAC;YAAE,OAAO,IAAI,CAAC;QAC5C,IAAI,kBAAkB,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,KAAK,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;YAAE,OAAO,IAAI,CAAC;QAC/E;;;;;WAKG;QACH,IAAI,kCAAkC,CAAC,GAAG,CAAC,KAAK,CAAC;YAAE,OAAO,IAAI,CAAC;QAC/D;;;;;;WAMG;QACH,OAAO,4BAA4B,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,IAAI,CAAC,wBAAwB,CAAC,KAAK,CAAC,CAAC;IACzF,CAAC;IAED;;;;;;OAMG;IACK,wBAAwB,CAAC,QAAgB;QAC/C,IAAI,CAAC,IAAI,CAAC,kBAAkB;YAAE,OAAO,KAAK,CAAC;QAC3C,IAAI,CAAC;YACH,MAAM,UAAU,GAAG,IAAI,CAAC,kBAAkB,CAAC,cAAc,CAAC,gCAAgC,CAAC,qBAAqB,CAAC,CAAC;YAClH,IAAI,UAAU,CAAC,UAAU,KAAK,KAAK;gBAAE,OAAO,KAAK,CAAC;YAClD,OAAO,UAAU,CAAC,OAAO,CAAC,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;QAClE,CAAC;QAAC,MAAM,CAAC;YACP,6FAA6F;YAC7F,wFAAwF;YACxF,OAAO,KAAK,CAAC;QACf,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,eAAe,CAAC,MAAgB,EAAE,MAAc;QAC5D,8FAA8F;QAC9F,2FAA2F;QAC3F,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;QAElE,IAAI,CAAC,QAAQ,CAAC,qCAAqC,EAAE;YACnD,MAAM;YACN,cAAc,EAAE,IAAI,CAAC,MAAM;SAC5B,CAAC,CAAC;QAEH,OAAO;YACL,IAAI;YACJ,QAAQ,EAAE;gBACR,cAAc,EAAE,IAAI,CAAC,MAAM;gBAC3B,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;aACpD;SACF,CAAC;IACJ,CAAC;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;OAoDG;IACK,KAAK,CAAC,mBAAmB,CAC/B,MAAgB,EAChB,MAAc,EACd,QAAgB,EAChB,OAAmC;QAEnC,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;QAChD,MAAM,OAAO,GAAG,OAAO,EAAE,uBAAuB,KAAK,IAAI;YACvD,CAAC,CAAC,IAAI,CAAC,4BAA4B,CAAC,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,MAAM,CAAC;YACpE,CAAC,CAAC,SAAS,CAAC;QAEd;;;;;;;;;;;;;WAaG;QACH,IAAI,CAAC,OAAO,IAAI,4BAA4B,CAAC,GAAG,CAAC,QAAQ,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC;YACzE,MAAM,IAAI,CAAC,YAAY,CAAC,UAAU,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,EAAE;gBAClE,sBAAsB,EAAE,2BAA2B,CAAC,qCAAqC;gBACzF,OAAO,EAAE;oBACP,wBAAwB;oBACxB,MAAM;oBACN,QAAQ;oBACR,UAAU,EAAE,MAAM,CAAC,MAAM;oBACzB,OAAO,EAAE,iHAAiH;iBAC3H;aACF,CAAC,CAAC;QACL,CAAC;QAED,IAAI,CAAC,OAAO,IAAI,OAAO,EAAE,uBAAuB,KAAK,IAAI,EAAE,CAAC;YAC1D,uFAAuF;YACvF,4FAA4F;YAC5F,2FAA2F;YAC3F,4CAA4C;YAC5C,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,4DAA4D,CAAC,CAAC;QAC3H,CAAC;QAED,IAAI,CAAC,OAAO,EAAE,CAAC;YACb,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,oCAAoC,CAAC,CAAC;QACnG,CAAC;QAED,IAAI,CAAC;YACH,MAAM,SAAS,GAAG,MAAM,IAAI,CAAC,kBAAmB,CAAC,mBAAmB,CAAC;gBACnE,QAAQ,EAAE,MAAM;gBAChB,QAAQ;gBACR,aAAa,EAAE,gCAAgC,CAAC,qBAAqB;gBACrE,MAAM;aACP,CAAC,CAAC;YAEH,IAAI,SAAS,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACrC,IAAI,CAAC,QAAQ,CAAC,gEAAgE,EAAE;oBAC9E,MAAM;oBACN,WAAW,EAAE,OAAO,CAAC,WAAW;oBAChC,KAAK,EAAE,SAAS,CAAC,KAAK;oBACtB,SAAS,EAAE,SAAS,CAAC,SAAS;oBAC9B,cAAc,EAAE,SAAS,CAAC,IAAI,CAAC,MAAM;iBACtC,CAAC,CAAC;gBACH,OAAO;oBACL,IAAI,EAAE,SAAS,CAAC,IAAI;oBACpB,QAAQ,EAAE;wBACR,SAAS,EAAE,SAAS,CAAC,SAAS;wBAC9B,cAAc,EAAE,SAAS,CAAC,IAAI,CAAC,MAAM;wBACrC,SAAS,EAAE,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;wBAC7D,cAAc,EAAE,UAAU;wBAC1B,qBAAqB,EAAE,OAAO,CAAC,WAAW;wBAC1C,eAAe,EAAE,SAAS,CAAC,KAAK;wBAChC,GAAG,CAAC,SAAS,CAAC,KAAK,CAAC,cAAc,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,sBAAsB,EAAE,SAAS,CAAC,KAAK,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;qBACpH;iBACF,CAAC;YACJ,CAAC;YAED,6FAA6F;YAC7F,8FAA8F;YAC9F,2FAA2F;YAC3F,+CAA+C;YAC/C,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,mFAAmF,EAAE;gBACpG,MAAM;gBACN,WAAW,EAAE,OAAO,CAAC,WAAW;gBAChC,SAAS,EAAE,SAAS,CAAC,SAAS;aAC/B,CAAC,CAAC;YACH,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,2CAA2C,CAAC,CAAC;QAC1G,CAAC;QAAC,OAAO,aAAa,EAAE,CAAC;YACvB,IACE,aAAa,YAAY,uBAAuB;mBAC7C,aAAa,CAAC,YAAY,KAAK,8BAA8B,CAAC,oBAAoB,EACrF,CAAC;gBACD,yCAAyC;gBACzC,IAAI,WAAiD,CAAC;gBACtD,IAAI,CAAC;oBACH,WAAW,GAAG,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,sCAAsC,CAAC,CAAC;gBAC5G,CAAC;gBAAC,MAAM,CAAC;oBACP,wFAAwF;oBACxF,qFAAqF;oBACrF,wFAAwF;oBACxF,qDAAqD;oBACrD,WAAW,GAAG,SAAS,CAAC;gBAC1B,CAAC;gBAED,IAAI,WAAW,IAAI,WAAW,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBACtD,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,4FAA4F,EAAE;wBAC7G,MAAM;wBACN,WAAW,EAAE,OAAO,CAAC,WAAW;wBAChC,UAAU,EAAE,aAAa,CAAC,UAAU;wBACpC,MAAM,EAAE,aAAa,CAAC,OAAO;wBAC7B,cAAc,EAAE,WAAW,CAAC,IAAI,CAAC,MAAM;qBACxC,CAAC,CAAC;oBACH,OAAO,WAAW,CAAC;gBACrB,CAAC;gBAED,IAAI,CAAC,MAAM,CAAC,IAAI,CACd,6IAA6I,EAC7I,EAAE,MAAM,EAAE,WAAW,EAAE,OAAO,CAAC,WAAW,EAAE,UAAU,EAAE,aAAa,CAAC,UAAU,EAAE,MAAM,EAAE,aAAa,CAAC,OAAO,EAAE,CAClH,CAAC;gBACF,wFAAwF;gBACxF,4FAA4F;gBAC5F,MAAM,aAAa,CAAC;YACtB,CAAC;YAED,0FAA0F;YAC1F,2FAA2F;YAC3F,sFAAsF;YACtF,yFAAyF;YACzF,IAAI,CAAC,QAAQ,CAAC,qFAAqF,EAAE;gBACnG,MAAM;gBACN,WAAW,EAAE,OAAO,CAAC,WAAW;gBAChC,MAAM,EAAE,aAAa,YAAY,KAAK,CAAC,CAAC,CAAC,aAAa,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC;aACvF,CAAC,CAAC;YACH,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,sDAAsD,CAAC,CAAC;QACrH,CAAC;IACH,CAAC;IAED;;;;;;;;;OASG;IACK,4BAA4B,CAClC,MAAc,EACd,QAAgB,EAChB,UAAkB;QAElB,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,CAAC;YAC7B,OAAO,SAAS,CAAC;QACnB,CAAC;QACD,IAAI,CAAC;YACH,MAAM,UAAU,GAAG,IAAI,CAAC,kBAAkB,CAAC,cAAc,CAAC,gCAAgC,CAAC,qBAAqB,CAAC,CAAC;YAClH,6FAA6F;YAC7F,0FAA0F;YAC1F,yFAAyF;YACzF,IAAI,UAAU,CAAC,UAAU,KAAK,KAAK,EAAE,CAAC;gBACpC,IAAI,CAAC,QAAQ,CAAC,sEAAsE,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,CAAC,MAAM,EAAE,CAAC,CAAC;gBAC7H,OAAO,SAAS,CAAC;YACnB,CAAC;YACD,IAAI,CAAC,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,UAAU,CAAC,OAAO,EAAE,QAAQ,EAAE,UAAU,CAAC,EAAE,CAAC;gBAChF,IAAI,CAAC,QAAQ,CAAC,8FAA8F,EAAE;oBAC5G,MAAM;oBACN,QAAQ;oBACR,UAAU;oBACV,gBAAgB,EAAE,UAAU,CAAC,OAAO,CAAC,gBAAgB;oBACrD,WAAW,EAAE,UAAU,CAAC,OAAO,CAAC,WAAW;iBAC5C,CAAC,CAAC;gBACH,OAAO,SAAS,CAAC;YACnB,CAAC;YACD,OAAO,UAAU,CAAC,OAAO,CAAC;QAC5B,CAAC;QAAC,OAAO,eAAe,EAAE,CAAC;YACzB,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,qGAAqG,EAAE;gBACtH,MAAM;gBACN,MAAM,EAAE,eAAe,YAAY,KAAK,CAAC,CAAC,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,eAAe,CAAC;aAC7F,CAAC,CAAC;YACH,OAAO,SAAS,CAAC;QACnB,CAAC;IACH,CAAC;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;OA4BG;IACH;;;;;;;;;;;;OAYG;IACK,KAAK,CAAC,cAAc,CAC1B,MAAc,EACd,MAAc,EACd,QAAgB,EAChB,MAAc;QAEd,IAAI,kCAAkC,CAAC,GAAG,CAAC,QAAQ,CAAC,WAAW,EAAE,CAAC,EAAE,CAAC;YACnE,OAAO,MAAM,IAAI,CAAC,yBAAyB,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,CAAC,WAAW,EAAE,EAAE,MAAM,CAAC,CAAC;QAC9F,CAAC;QACD,OAAO,MAAM,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC;IAC3D,CAAC;IAED;;;;;;;;;;;;;OAaG;IACK,KAAK,CAAC,yBAAyB,CACrC,MAAc,EACd,MAAc,EACd,QAAgB,EAChB,MAAc;QAEd,MAAM,SAAS,GAAG,MAAM,yBAAyB,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;QAEpE,IAAI,CAAC,QAAQ,CAAC,wEAAwE,EAAE;YACtF,MAAM;YACN,QAAQ;YACR,MAAM;YACN,SAAS,EAAE,SAAS,CAAC,SAAS;YAC9B,cAAc,EAAE,SAAS,CAAC,IAAI,CAAC,MAAM;SACtC,CAAC,CAAC;QAEH,OAAO;YACL,IAAI,EAAE,SAAS,CAAC,IAAI;YACpB,QAAQ,EAAE;gBACR,SAAS,EAAE,SAAS,CAAC,SAAS;gBAC9B,cAAc,EAAE,SAAS,CAAC,IAAI,CAAC,MAAM;gBACrC,SAAS,EAAE,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;gBAC7D,cAAc,EAAE,eAAe;aAChC;SACF,CAAC;IACJ,CAAC;IAEO,KAAK,CAAC,cAAc,CAAC,MAAc,EAAE,MAAc,EAAE,MAAc;QACzE,MAAM,KAAK,GAAG,MAAM,MAAM,CAAC,iCAAiC,CAAC,CAAC;QAE9D,MAAM,WAAW,GAAG,KAAK,CAAC,WAAW,CAAC;YACpC,kFAAkF;YAClF,yFAAyF;YACzF,sFAAsF;YACtF,0FAA0F;YAC1F,yFAAyF;YACzF,eAAe;YACf,IAAI,EAAE,IAAI,UAAU,CAAC,MAAM,CAAC;YAC5B,uFAAuF;YACvF,sEAAsE;YACtE,EAAE;YACF,4FAA4F;YAC5F,qFAAqF;YACrF,8EAA8E;YAC9E,EAAE;YACF,sFAAsF;YACtF,yFAAyF;YACzF,0EAA0E;YAC1E,EAAE;YACF,2FAA2F;YAC3F,sFAAsF;YACtF,4FAA4F;YAC5F,0CAA0C;YAC1C,cAAc,EAAE,KAAK;YACrB,mBAAmB,EAAE,2BAA2B,EAAE;SACnD,CAAC,CAAC;QAEH,IAAI,CAAC;YACH,MAAM,WAAW,GAAG,MAAM,WAAW,CAAC,OAAO,CAAC;YAC9C,MAAM,SAAS,GAAa,EAAE,CAAC;YAE/B,KAAK,IAAI,UAAU,GAAG,CAAC,EAAE,UAAU,IAAI,WAAW,CAAC,QAAQ,EAAE,UAAU,EAAE,EAAE,CAAC;gBAC1E,MAAM,IAAI,GAAG,MAAM,WAAW,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;gBACnD,IAAI,CAAC;oBACH,MAAM,WAAW,GAAG,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;oBAChD,qFAAqF;oBACrF,EAAE;oBACF,kFAAkF;oBAClF,yFAAyF;oBACzF,wFAAwF;oBACxF,sFAAsF;oBACtF,gEAAgE;oBAChE,EAAE;oBACF,uFAAuF;oBACvF,uFAAuF;oBACvF,8DAA8D;oBAC9D,IAAI,QAAQ,GAAG,EAAE,CAAC;oBAClB,KAAK,MAAM,IAAI,IAAI,WAAW,CAAC,KAAK,EAAE,CAAC;wBACrC,IAAI,CAAC,CAAC,KAAK,IAAI,IAAI,CAAC;4BAAE,SAAS;wBAC/B,QAAQ,IAAI,IAAI,CAAC,GAAG,CAAC;wBACrB,IAAI,IAAI,CAAC,MAAM;4BAAE,QAAQ,IAAI,IAAI,CAAC;oBACpC,CAAC;oBACD,SAAS,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;gBAC3B,CAAC;wBAAS,CAAC;oBACT,IAAI,CAAC,OAAO,EAAE,CAAC;gBACjB,CAAC;YACH,CAAC;YAED,MAAM,IAAI,GAAG,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAElC,IAAI,CAAC,QAAQ,CAAC,uDAAuD,EAAE;gBACrE,MAAM;gBACN,SAAS,EAAE,WAAW,CAAC,QAAQ;gBAC/B,cAAc,EAAE,IAAI,CAAC,MAAM;gBAC3B,MAAM;aACP,CAAC,CAAC;YAEH,OAAO;gBACL,IAAI;gBACJ,QAAQ,EAAE;oBACR,SAAS,EAAE,WAAW,CAAC,QAAQ;oBAC/B,cAAc,EAAE,IAAI,CAAC,MAAM;oBAC3B,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM;oBACnD,wFAAwF;oBACxF,oFAAoF;oBACpF,cAAc,EAAE,eAAe;oBAC/B,oBAAoB,EAAE,MAAM;iBAC7B;aACF,CAAC;QACJ,CAAC;gBAAS,CAAC;YACT,8FAA8F;YAC9F,oFAAoF;YACpF,EAAE;YACF,8FAA8F;YAC9F,wFAAwF;YACxF,6FAA6F;YAC7F,6DAA6D;YAC7D,MAAM,WAAW,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;QACrD,CAAC;IACH,CAAC;IAED,qFAAqF;IAC7E,KAAK,CAAC,aAAa,CAAC,MAAgB;QAC1C,MAAM,MAAM,GAAa,EAAE,CAAC;QAC5B,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;YACjC,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;QACnE,CAAC;QACD,OAAO,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IAC/B,CAAC;CACF,CAAA;AAjnBY,gCAAgC;IAD5C,UAAU,EAAE;IAKR,WAAA,MAAM,CAAC,kBAAkB,CAAC,MAAM,CAAC,CAAA;IAEjC,WAAA,MAAM,CAAC,kBAAkB,CAAC,mBAAmB,CAAC,CAAA;IAE9C,WAAA,MAAM,CAAC,kBAAkB,CAAC,iBAAiB,CAAC,CAAA;IAG5C,WAAA,MAAM,CAAC,kBAAkB,CAAC,mBAAmB,CAAC,CAAA;IAQ9C,WAAA,MAAM,CAAC,kBAAkB,CAAC,yBAAyB,CAAC,CAAA;IACpD,WAAA,QAAQ,EAAE,CAAA;+CAboB,0BAA0B;QAErB,+BAA+B;QAGpC,mBAAmB;QASZ,gCAAgC;GArB7D,gCAAgC,CAinB5C","sourcesContent":["/**\n * File Text Extraction Backend Service\n *\n * Extracts text content from files. Retained through the Weaviate-lane demolition (2026-08-26):\n * this is the file-side TEXT PROVIDER the resource-native RAG capability reuses in Phase 1 of\n * `.claude/plans/resource-rag-capability.md` (extraction knobs like extractText/ocr stay file-side).\n *\n * Supported formats (first slice):\n * - Plain text (text/plain, text/markdown, text/csv, application/json, etc.)\n * - PDF (via `pdfjs-dist` — dynamically imported, but an INSTALLED dependency)\n *\n * ## Why `pdfjs-dist` and not `pdf-parse` (decided 2026-08-27)\n *\n * The header used to advertise PDF support through `pdf-parse`, which was in NO manifest and NO\n * lockfile anywhere in the repository. So every PDF that reached a `.ragSource()` field threw\n * `'pdf-parse is required for PDF text extraction'` — and because the RAG sweep treats a throw as a\n * failed ingestion and re-visits hourly, that was a permanent retry loop, not a one-off failure.\n * The advertised capability had never worked. Both halves of that are now fixed: the library is a\n * real dependency, and the sweep no longer strands on an unextractable file.\n *\n * `pdf-parse` was reconsidered rather than assumed obsolete — its modern release (2.4.5, Oct 2025)\n * is maintained, contrary to the reputation of the long-abandoned 1.1.x line. It was still rejected,\n * on dependency shape rather than hygiene: it requires `@napi-rs/canvas` (a NATIVE addon, per-platform\n * binaries, and this repository's own toolchain baseline records what a stale C++ toolchain costs)\n * and it vendors `pdfjs-dist@5.4.296` — a SECOND copy at a different major from the `^6.2.108` this\n * repository already pins centrally. `pdfjs-dist` used directly has zero required dependencies\n * (`@napi-rs/canvas` is OPTIONAL there and is needed only for RENDERING, which text extraction never\n * does), is Mozilla-maintained, and matches the `^6.2.108` the root `pnpm.overrides` already pins\n * for the PDF viewer preset. One library, one major, no native build.\n *\n * (The `pdf-parse` dependency facts above were read from the npm registry on 2026-08-27 — `npm view\n * pdf-parse dependencies` reports `pdfjs-dist@5.4.296` and `@napi-rs/canvas@0.1.80`. They are not\n * checkable from inside this repository, so the provenance is recorded rather than assumed.)\n *\n * ## The version constraint is a RANGE, not the usual `*`\n *\n * Most dependencies here are declared `*` because the root `pnpm.overrides` is the single version\n * authority. That works inside the monorepo and NOT for a generated application, which inherits no\n * root override — and `pdfjs-dist` is not in `propagated-engine-overrides.json`. Under `*` an\n * application would happily satisfy this import with the `4.10.38` its PDF viewer already pulls in\n * transitively, against which this file's API assumptions are false (`loadingTask.destroy`, the\n * absent `exports` map that makes both the deep `legacy/build` specifier and the `package.json`\n * resolution legal). `^6.2.108` states what the CODE requires. Precedent: `@react-pdf/renderer`\n * carries a real range in this same manifest for the same reason.\n *\n * ## Provider-backed extraction (OCR) — the selection seam\n *\n * A PDF is served by the application's `EngineCapability.AI_DOCUMENT_EXTRACTION` provider when one\n * is configured and can handle the bytes, and by the local library otherwise. `extractFromDocument`\n * owns that decision and documents the full precedence, including the branch that protects an\n * existing RAG corpus from a provider outage. Plain text never touches a provider.\n *\n * ## Office / OpenDocument — six served LOCALLY, three provider-only (#434, closed 2026-09-09)\n *\n * The framework's `Documents` media group admits Word, Excel, PowerPoint and the OpenDocument\n * trio, and until 2026-09-08 the extractor supported NONE of them: a corpus field using the group\n * accepted uploads that could never contribute a chunk under any configuration, skipped at debug.\n *\n * The nine types are THREE technical problems, and reading them as one is what produced this\n * header's earlier claim that a local library would cost \"three to six\" dependencies:\n *\n * - **OOXML** (`.docx` `.xlsx` `.pptx`) and **OpenDocument** (`.odt` `.ods` `.odp`) are ZIP+XML —\n *   the SAME technique, so ONE module reads all six, on two dependencies rather than six\n *   (`fflate` to unzip, `fast-xml-parser` to read). See\n *   `office-document-text-extraction.backend.utils.ts`.\n * - **Legacy binary Office** (`.doc` `.xls` `.ppt`) are OLE2 compound files from a different era.\n *   A reader for them is a project, not a dependency, so they stay PROVIDER-ONLY.\n *\n * So the six take exactly the ladder PDF already takes, with the office extractor in the rung\n * `pdfjs-dist` occupies: a configured provider is preferred when the field opted into it, the\n * local library serves everything else, and a deployment with NO provider can finally read the\n * documents its media group accepts. The three remaining types keep the conditional support and\n * the precise refusal, because for them the provider genuinely is the only path.\n *\n * Future:\n * - Legacy binary Office (`.doc` / `.xls` / `.ppt`) locally, if an OLE2 reader ever earns its place\n * - Audio/video transcript extraction\n *\n * The service operates against the storage provider contract — it downloads\n * the file via `FileStorageRouterBackendService`, never touching raw S3/local\n * paths directly.\n */\n\nimport { injectable, inject, optional } from 'inversify';\nimport { Readable } from 'stream';\nimport { createRequire } from 'node:module';\nimport { dirname, join } from 'node:path';\nimport { ErrorType, ErrorCustomMessageReference, type File } from '@wildo-ai/saas-models';\nimport {\n  extractOfficeDocumentText,\n  LEGACY_BINARY_OFFICE_MIMES,\n  LOCALLY_EXTRACTABLE_DOCUMENT_MIMES,\n  OfficeDocumentExtractionError,\n} from '~wildo-backend-lib/files/utils/office-document-text-extraction.backend.utils';\nimport { SAAS_SERVICE_TYPES } from '~wildo-backend-lib/inject/models/service-types';\nimport type { LoggerService } from '~wildo-backend-lib/monitoring/logger/services/logger.service';\nimport { ErrorBuilderBackendService } from '~wildo-backend-lib/errors/services/error-builder.backend.service';\nimport { ExecutionContext } from '~wildo-backend-lib/runtime/models/execution-context.backend.schemas';\nimport { FileStorageRouterBackendService } from '~wildo-backend-lib/files/storage/file-storage-router.backend.service';\nimport { FilesBackendService } from '~wildo-backend-lib/files/services/files.backend.service';\nimport {\n  DocumentExtractionBackendService,\n  DocumentExtractionError,\n  DocumentExtractionFailureClass,\n  type DocumentExtractionResolvedBinding,\n} from '~wildo-backend-lib/external-providers/services/document-extraction.backend.service';\nimport { DocumentExtraction_Applicability } from '@wildo-ai/external-connectors-models';\n\n/**\n * Per-call extraction options — today, the caller's answer to \"may this document be sent to an\n * external extraction provider?\".\n *\n * The answer belongs to the CALLER because it is a per-FIELD declaration\n * (`z_file({ textExtraction: { ocr } })`) and this service is handed a `File`, which knows nothing\n * about the field it hangs off. `RagIngestionBackendService` is the layer that holds both, so it\n * is the layer that translates one into the other.\n */\nexport interface FileTextExtractionOptions {\n  /**\n   * Whether provider-backed extraction (OCR) is permitted for this document.\n   *\n   * **DEFAULT FALSE, deliberately, and it is a spending decision rather than a caution.** OCR is\n   * billed per page by an external provider, and `FileTextExtractionOptions.ocr` on the schema is\n   * documented to application developers as off-by-default for exactly that reason. Defaulting\n   * this to true would mean that merely CONFIGURING a provider silently began charging per page\n   * for every document in every corpus — including fields whose author wrote `ocr: false` and\n   * explained why. Wonder Todos' `knowledge-documents` field is precisely that case.\n   *\n   * When false, extraction uses the local text-layer read and nothing leaves the deployment.\n   */\n  allowProviderExtraction?: boolean;\n}\n\nexport interface FileTextExtractionResult {\n  text: string;\n  language?: string;\n  confidence?: number;\n  metadata?: {\n    pageCount?: number;\n    wordCount?: number;\n    characterCount?: number;\n    [key: string]: unknown;\n  };\n}\n\nconst TEXT_MIME_PREFIXES = ['text/'];\nconst TEXT_MIME_EXACT = new Set([\n  'application/json',\n  'application/xml',\n  'application/xhtml+xml',\n  'application/javascript',\n  'application/typescript',\n  'application/x-yaml',\n  'application/yaml',\n  'application/x-sh',\n  'application/graphql',\n  'application/toml',\n]);\nconst PDF_MIME = 'application/pdf';\n\n/**\n * The three types that STILL need a provider — pre-2007 binary Office (`.doc` / `.xls` / `.ppt`).\n *\n * These are OLE2 compound files, not ZIP archives, so the local ZIP+XML extractor that serves the\n * other six cannot read them and a reader for them is a project rather than a dependency. Support\n * is therefore CONDITIONAL, and {@link isExtractionSupported} says so by ASKING the provider rather\n * than asserting a capability the deployment may not have.\n *\n * Why conditional rather than unconditional-and-fail-later: RAG ingestion treats an unsupported\n * MIME as \"no corpus from this file\", logged at DEBUG and never an error. Declaring these supported\n * everywhere would turn a configuration ABSENCE into an ERROR log on every sweep run, for every\n * application without an extraction provider — noise that says nothing an operator can act on.\n *\n * The six ZIP+XML types are no longer here, and that is the whole of #434's second half: they are\n * served by {@link LOCALLY_EXTRACTABLE_DOCUMENT_MIMES} with no provider required.\n */\nconst PROVIDER_ONLY_DOCUMENT_MIMES: ReadonlySet<string> = LEGACY_BINARY_OFFICE_MIMES;\n\nconst moduleRequire = createRequire(import.meta.url);\n\n/**\n * Absolute path to `pdfjs-dist`'s bundled standard-font data, resolved from the INSTALLED package\n * rather than assumed relative to this file — `saas-backend-lib` runs from `dist/` in every\n * deployed process, so a path relative to the source tree would be wrong everywhere that matters.\n *\n * Without this, pdf.js emits `Warning: UnknownErrorException: Ensure that the 'standardFontDataUrl'\n * API parameter is provided` for any document using the 14 PDF standard fonts (Helvetica, Times,\n * Courier …) — the family ordinary word processors and report generators emit.\n *\n * MEASURED 2026-08-27, because the obvious claim about this is wrong: removing this parameter does\n * NOT break extraction of the committed Helvetica fixture. Text extraction reads the string\n * operands of the content stream and maps them through the font's ENCODING, which pdf.js can build\n * for a standard font without loading its font PROGRAM. So this is supplied to silence a real\n * warning and to keep glyph mapping correct for documents whose encoding genuinely requires the\n * font program — NOT because the common case depends on it. No fixture currently pins a document\n * that would fail without it; do not claim one does.\n *\n * Resolved lazily and memoized: `require.resolve` touches the filesystem, and extraction runs per\n * file inside a sweep that can process many files in one pass.\n */\nlet cachedStandardFontDataPath: string | undefined;\nfunction resolveStandardFontDataPath(): string {\n  if (cachedStandardFontDataPath === undefined) {\n    // The trailing separator is REQUIRED — pdf.js concatenates the font filename onto this value.\n    cachedStandardFontDataPath = join(dirname(moduleRequire.resolve('pdfjs-dist/package.json')), 'standard_fonts') + '/';\n  }\n  return cachedStandardFontDataPath;\n}\n\n@injectable()\nexport class FileTextExtractionBackendService {\n  protected logDebug: (message: string, context?: Record<string, unknown>) => void;\n\n  constructor(\n    @inject(SAAS_SERVICE_TYPES.Logger)\n    private readonly logger: LoggerService,\n    @inject(SAAS_SERVICE_TYPES.ErrorBuilderService)\n    private readonly errorBuilder: ErrorBuilderBackendService,\n    @inject(SAAS_SERVICE_TYPES.FileStorageRouter)\n    private readonly fileStorageRouter: FileStorageRouterBackendService,\n    // Appended LAST deliberately — DI constructor positions are order-sensitive.\n    @inject(SAAS_SERVICE_TYPES.FilesBackendService)\n    private readonly filesService: FilesBackendService,\n    // OPTIONAL, and that is not defensiveness. This service is constructed directly in several\n    // tests and integration harnesses with a fixed positional argument list; a required dependency\n    // appended here would break every one of them, and — worse — a harness that passed `undefined`\n    // to satisfy the arity would look wired while being dead. Optional makes \"no provider\" the\n    // explicit, tested path it genuinely is: an application with no `AI_DOCUMENT_EXTRACTION`\n    // selection behaves identically to one whose container never bound this service.\n    @inject(SAAS_SERVICE_TYPES.DocumentExtractionService)\n    @optional()\n    private readonly documentExtraction?: DocumentExtractionBackendService,\n  ) {\n    this.logDebug = this.logger.debugByType('FileTextExtractionBackendService');\n    this.logger.setDebugTypeLogging('FileTextExtractionBackendService', false);\n  }\n\n  /**\n   * Extract text from a file by downloading it through the storage provider.\n   */\n  async extractText(\n    file: File,\n    executionContext: ExecutionContext<any>,\n    options?: FileTextExtractionOptions,\n  ): Promise<FileTextExtractionResult> {\n    const mimeType = file.mimeType?.toLowerCase() ?? '';\n\n    if (!this.isExtractionSupported(mimeType)) {\n      throw this.errorBuilder.buildError(\n        ErrorType.VALIDATION,\n        executionContext,\n        {\n          customMessageReference: ErrorCustomMessageReference.FILE_TEXT_EXTRACTION_UNSUPPORTED_TYPE,\n          context: {\n            // diagnostic (log-only)\n            fileId: file.fileId,\n            mimeType,\n          },\n        },\n      );\n    }\n\n    // Resolve through FilesBackendService so a LEGACY row that persisted no\n    // `storageProvider` falls back to the default provider instead of failing\n    // extraction. Reimplementing the lookup here is what made this one of five\n    // behaviours for one input.\n    //\n    // Unlike the scan pipeline, the throw is propagated: this runs on a synchronous\n    // request path with no terminal status to strand, so failing loudly is correct.\n    // The error type moves CONFIGURATION -> INTERNAL, matching the reference — an\n    // unregistered persisted id is an engine wiring fault, not an app misconfiguration.\n    const provider = this.filesService.getStorageProviderForFile(file, executionContext);\n\n    const downloadResult = await provider.download(\n      { fileId: file.fileId, storageProvider: file.storageProvider, storageRef: file.storageRef },\n      executionContext,\n    );\n\n    try {\n      try {\n        if (\n          mimeType === PDF_MIME\n          || LOCALLY_EXTRACTABLE_DOCUMENT_MIMES.has(mimeType)\n          || PROVIDER_ONLY_DOCUMENT_MIMES.has(mimeType)\n        ) {\n          return await this.extractFromDocument(downloadResult.stream, file.fileId, mimeType, options);\n        }\n        return await this.extractFromText(downloadResult.stream, file.fileId);\n      } catch (parseError) {\n        // A provider fault that could not be covered by the local library reaches here as a\n        // `DocumentExtractionError` of class PROVIDER_UNAVAILABLE, and it MUST NOT be classified\n        // as unreadable bytes — see `extractFromDocument` for the full reasoning. Rethrowing it\n        // untouched puts it in the same bucket as a storage outage, which is exactly right: both\n        // mean \"we could not read this document RIGHT NOW\", and both must fail the ingestion so\n        // the existing corpus survives.\n        if (\n          parseError instanceof DocumentExtractionError\n          && parseError.failureClass === DocumentExtractionFailureClass.PROVIDER_UNAVAILABLE\n        ) {\n          throw parseError;\n        }\n        // CLASSIFY the failure, because callers must be able to tell two very different things\n        // apart and a raw pdf.js exception does not let them.\n        //\n        // Everything ABOVE this point — resolving the storage provider, downloading the blob — can\n        // fail for INFRASTRUCTURE reasons: an unregistered provider id (a wiring fault), or an\n        // unreachable object store (transient). Those errors propagate untouched, and must, because\n        // a batch consumer that treats them as \"this file is unreadable\" would draw a permanent\n        // conclusion from a temporary outage.\n        //\n        // Everything INSIDE this try is parsing bytes we already hold. A failure here is a property\n        // of the document — corrupt, truncated, password-protected — and re-reading it later cannot\n        // help. Only that class is safe for a sweep to skip.\n        //\n        // The distinction is not cosmetic: `RagIngestionBackendService` deletes a field's chunk set\n        // when a pass yields no text, so mislabelling a storage outage as unreadable bytes would\n        // destroy a live corpus.\n        throw this.errorBuilder.buildError(ErrorType.VALIDATION, executionContext, {\n          customMessageReference: ErrorCustomMessageReference.FILE_TEXT_EXTRACTION_FAILED,\n          context: {\n            // diagnostic (log-only)\n            fileId: file.fileId,\n            mimeType,\n            cause: parseError instanceof Error ? parseError.message : String(parseError),\n          },\n        });\n      }\n    } finally {\n      if (!downloadResult.stream.destroyed) {\n        downloadResult.stream.destroy();\n      }\n    }\n  }\n\n  /**\n   * Check if a MIME type supports text extraction.\n   */\n  isExtractionSupported(mimeType: string): boolean {\n    const lower = mimeType.toLowerCase();\n    if (lower === PDF_MIME) return true;\n    if (TEXT_MIME_EXACT.has(lower)) return true;\n    if (TEXT_MIME_PREFIXES.some((prefix) => lower.startsWith(prefix))) return true;\n    /*\n     * OOXML + OpenDocument: UNCONDITIONALLY supported (#434). A local ZIP+XML extractor serves all\n     * six, so the answer no longer depends on a deployment's provider selection — which is the\n     * point of the local library. A provider, where one is configured and the field opted in, is\n     * still preferred inside `extractFromDocument`; it is no longer the only path.\n     */\n    if (LOCALLY_EXTRACTABLE_DOCUMENT_MIMES.has(lower)) return true;\n    /*\n     * Legacy binary Office: supported IF AND ONLY IF a configured provider declares the type. Byte\n     * length is unknown at this question — it is asked before the file is fetched — so this answers\n     * the MIME half only; `resolveUsableProviderBinding` re-asks with the real length and falls\n     * through if the document is too large, which is the same shape as a PDF exceeding\n     * `maxDocumentBytes`.\n     */\n    return PROVIDER_ONLY_DOCUMENT_MIMES.has(lower) && this.providerDeclaresMimeType(lower);\n  }\n\n  /**\n   * Whether the configured document-extraction provider declares this MIME type.\n   *\n   * Deliberately narrower than {@link resolveUsableProviderBinding}: no byte length, no logging, and\n   * every failure answers `false`. It is a capability QUESTION asked on a hot path (once per file\n   * per sweep), not the resolution that precedes an actual extraction.\n   */\n  private providerDeclaresMimeType(mimeType: string): boolean {\n    if (!this.documentExtraction) return false;\n    try {\n      const resolution = this.documentExtraction.resolveBinding(DocumentExtraction_Applicability.DOCUMENT_TEXT_DEFAULT);\n      if (resolution.configured === false) return false;\n      return resolution.binding.supportedMimeTypes.includes(mimeType);\n    } catch {\n      // A misconfigured provider is not a supported type. The loud version of this is logged where\n      // extraction is actually attempted; answering `false` here keeps the corpus skip quiet.\n      return false;\n    }\n  }\n\n  private async extractFromText(stream: Readable, fileId: string): Promise<FileTextExtractionResult> {\n    // Plain text keeps its ZERO-dependency path deliberately: routing text/* through a PDF parser\n    // or a remote provider would add cost and an egress for bytes that are already the answer.\n    const text = (await this.collectStream(stream)).toString('utf-8');\n\n    this.logDebug('Extracted text from plain text file', {\n      fileId,\n      characterCount: text.length,\n    });\n\n    return {\n      text,\n      metadata: {\n        characterCount: text.length,\n        wordCount: text.split(/\\s+/).filter(Boolean).length,\n      },\n    };\n  }\n\n  /**\n   * THE SELECTION SEAM — provider first when one is configured and usable, local library\n   * otherwise, and a deliberate answer for every way both can fail.\n   *\n   * ## The precedence, and why each rung is where it is\n   *\n   * 0. **The FIELD did not opt into provider extraction** → local library, with no provider\n   *    resolution attempted at all. `z_file({ textExtraction: { ocr } })` is documented to\n   *    application developers as off-by-default because OCR is an external per-page cost, so\n   *    \"an OCR provider is configured somewhere in this application\" must never be sufficient to\n   *    start spending on a field whose author declared `ocr: false`. Two conditions, both\n   *    required — the same composition as `.ragSource()` + `AI_EMBEDDINGS`.\n   * 1. **No provider configured** → local library. Identical to the behaviour before this seam\n   *    existed, and it is the default: an application that ingests only born-digital PDFs never\n   *    needs OCR and never pays for an egress.\n   * 2. **Provider configured but cannot serve THESE bytes** (MIME outside the model's catalog row,\n   *    or larger than the effective ceiling) → local library. Checked with `canServe()`, which is\n   *    a pure catalog lookup: learning this from a rejected request would cost an egress of the\n   *    application's document to discover something already known.\n   * 3. **Provider returns text** → provider wins. This is the whole point — a scanned contract has\n   *    no text layer, so this is the only rung that can produce anything for it.\n   * 4. **Provider returns EMPTY text** → try the local library anyway. Cheap, and the two read\n   *    genuinely different things: an OCR model can return nothing for a page whose text layer\n   *    pdf.js reads perfectly well (a vector-drawn document, say). Preferring \"some text\" over\n   *    \"no text\" needs no further justification.\n   * 5. **Provider says the DOCUMENT is unreadable** (`UNREADABLE_DOCUMENT`) → try the local\n   *    library, then classify as unreadable if it too yields nothing. Permanent by nature, so the\n   *    sweep should skip rather than retry.\n   * 6. **Provider was UNREACHABLE** (`PROVIDER_UNAVAILABLE` — network, 5xx, 429, bad credential,\n   *    timeout) → try the local library. If it yields text, serve that, degraded but correct. If\n   *    it yields NOTHING, **rethrow the provider fault** rather than reporting empty text.\n   *\n   * Rung 6 is the one that matters, and it is the reason this method is not four lines long.\n   *\n   * `RagIngestionBackendService` deletes a field's entire chunk set when an ingestion pass yields\n   * no text. So for a SCANNED document — one that only the provider can read — a provider outage\n   * plus a \"return empty text\" response would silently destroy a live, already-embedded corpus and\n   * report success, and the next sweep after the outage would rebuild it. That is data loss caused\n   * by a transient network fault, which is not an acceptable trade for a tidier code path.\n   * Rethrowing makes the ingestion FAIL, which leaves the existing chunks untouched and lets the\n   * sweep retry next hour — the treatment that already exists for a storage outage, reused because\n   * the two situations are the same situation.\n   *\n   * The distinction is only expressible because the provider service CLASSIFIES its failures; a\n   * seam that caught a bare `Error` could not tell rung 5 from rung 6, and would have to pick one\n   * wrong answer for the other.\n   *\n   * ## What this method never does\n   *\n   * It never routes `text/*` anywhere near a provider — plain text keeps its zero-dependency path,\n   * because the bytes already ARE the answer and sending them would be a pointless egress of\n   * application content. That gate lives in `extractText` above, not here.\n   */\n  private async extractFromDocument(\n    stream: Readable,\n    fileId: string,\n    mimeType: string,\n    options?: FileTextExtractionOptions,\n  ): Promise<FileTextExtractionResult> {\n    const buffer = await this.collectStream(stream);\n    const binding = options?.allowProviderExtraction === true\n      ? this.resolveUsableProviderBinding(fileId, mimeType, buffer.length)\n      : undefined;\n\n    /*\n     * LEGACY BINARY Office only (#434). These are OLE2 compound files, so neither local library\n     * below can serve them — handing a `.doc` to pdf.js or to the ZIP+XML extractor produces an\n     * unreadable-bytes classification that blames the document for a capability the deployment does\n     * not have.\n     *\n     * The six ZIP+XML types deliberately do NOT reach this branch any more: they have a local\n     * library, so a missing binding degrades them to it rather than refusing.\n     *\n     * `isExtractionSupported` already asked the provider whether it declares this MIME, so reaching\n     * here without a binding means the answer changed between the two questions (the byte length\n     * exceeded `maxDocumentBytes`, or the provider became unresolvable). Refusing names that\n     * precisely instead of mislabelling the file.\n     */\n    if (!binding && PROVIDER_ONLY_DOCUMENT_MIMES.has(mimeType.toLowerCase())) {\n      throw this.errorBuilder.buildError(ErrorType.VALIDATION, undefined, {\n        customMessageReference: ErrorCustomMessageReference.FILE_TEXT_EXTRACTION_UNSUPPORTED_TYPE,\n        context: {\n          // diagnostic (log-only)\n          fileId,\n          mimeType,\n          byteLength: buffer.length,\n          message: 'This document type is extractable only through a configured document-extraction provider, and none can serve it',\n        },\n      });\n    }\n\n    if (!binding && options?.allowProviderExtraction !== true) {\n      // Rung 0 — the field did not opt in. Not a fallback and not a degradation: this is the\n      // declared behaviour of a field whose author left `ocr` off (or wrote `ocr: false`), and it\n      // is where the overwhelming majority of documents belong, because a born-digital PDF gains\n      // nothing from OCR and would cost per page.\n      return await this.extractLocally(buffer, fileId, mimeType, 'local-library (field did not opt into provider extraction)');\n    }\n\n    if (!binding) {\n      return await this.extractLocally(buffer, fileId, mimeType, 'local-library (no usable provider)');\n    }\n\n    try {\n      const extracted = await this.documentExtraction!.extractDocumentText({\n        document: buffer,\n        mimeType,\n        applicability: DocumentExtraction_Applicability.DOCUMENT_TEXT_DEFAULT,\n        fileId,\n      });\n\n      if (extracted.text.trim().length > 0) {\n        this.logDebug('Extracted text via the configured document-extraction provider', {\n          fileId,\n          providerRef: binding.providerRef,\n          model: extracted.model,\n          pageCount: extracted.pageCount,\n          characterCount: extracted.text.length,\n        });\n        return {\n          text: extracted.text,\n          metadata: {\n            pageCount: extracted.pageCount,\n            characterCount: extracted.text.length,\n            wordCount: extracted.text.split(/\\s+/).filter(Boolean).length,\n            extractionPath: 'provider',\n            extractionProviderRef: binding.providerRef,\n            extractionModel: extracted.model,\n            ...(extracted.usage.pagesProcessed !== undefined ? { providerPagesProcessed: extracted.usage.pagesProcessed } : {}),\n          },\n        };\n      }\n\n      // Rung 4. WARN rather than debug: a provider that read the document and found no text at all\n      // is worth noticing — it is either a genuinely blank document or a model that failed quietly,\n      // and this service's debug channel is OFF by default, which is how the original \"every PDF\n      // throws\" defect stayed invisible for so long.\n      this.logger.warn('Document-extraction provider returned no text — falling back to the local library', {\n        fileId,\n        providerRef: binding.providerRef,\n        pageCount: extracted.pageCount,\n      });\n      return await this.extractLocally(buffer, fileId, mimeType, 'local-library (provider returned no text)');\n    } catch (providerError) {\n      if (\n        providerError instanceof DocumentExtractionError\n        && providerError.failureClass === DocumentExtractionFailureClass.PROVIDER_UNAVAILABLE\n      ) {\n        // Rung 6 — the corpus-protecting branch.\n        let localResult: FileTextExtractionResult | undefined;\n        try {\n          localResult = await this.extractLocally(buffer, fileId, mimeType, 'local-library (provider unavailable)');\n        } catch {\n          // The local library ALSO failed. Deliberately swallowed: the provider fault is the more\n          // informative of the two and is the one whose class the caller branches on, so it is\n          // what gets rethrown below. Losing a pdf.js parse message here costs a log line; losing\n          // the provider's classification would cost a corpus.\n          localResult = undefined;\n        }\n\n        if (localResult && localResult.text.trim().length > 0) {\n          this.logger.warn('Document-extraction provider was unavailable — served this document from the local library', {\n            fileId,\n            providerRef: binding.providerRef,\n            httpStatus: providerError.httpStatus,\n            reason: providerError.message,\n            characterCount: localResult.text.length,\n          });\n          return localResult;\n        }\n\n        this.logger.warn(\n          'Document-extraction provider was unavailable and the local library produced no text — failing the extraction to protect any existing corpus',\n          { fileId, providerRef: binding.providerRef, httpStatus: providerError.httpStatus, reason: providerError.message },\n        );\n        // NOT classified as unreadable bytes. `extractText`'s catch rethrows this untouched, so\n        // the RAG sweep reports FAILED and leaves existing chunks alone — see this method's header.\n        throw providerError;\n      }\n\n      // Rung 5, plus any non-classified throw (a CONFIGURATION error from resolveBinding, say).\n      // Both are treated the same way here on purpose: try the local library, and if it produces\n      // nothing let the failure fall through to `extractText`'s classifier. A configuration\n      // mistake reaching this point has already been logged as itself by the provider service.\n      this.logDebug('Document-extraction provider could not read the document — trying the local library', {\n        fileId,\n        providerRef: binding.providerRef,\n        reason: providerError instanceof Error ? providerError.message : String(providerError),\n      });\n      return await this.extractLocally(buffer, fileId, mimeType, 'local-library (provider could not read the document)');\n    }\n  }\n\n  /**\n   * The usable provider binding for this document, or `undefined` when the local library should\n   * serve — rungs 1 and 2 of the precedence in {@link extractFromDocument}.\n   *\n   * Never throws. A CONFIGURATION fault here (a provider selected but declaring no contract) is\n   * logged and treated as \"no provider\": extraction is an ENHANCEMENT over a local path that\n   * works, so an operator's provider mistake must not take down file extraction for documents the\n   * local library reads perfectly well. The provider service still throws that error loudly on the\n   * `extractDocumentText` path, so the mistake is not silenced — only prevented from cascading.\n   */\n  private resolveUsableProviderBinding(\n    fileId: string,\n    mimeType: string,\n    byteLength: number,\n  ): DocumentExtractionResolvedBinding | undefined {\n    if (!this.documentExtraction) {\n      return undefined;\n    }\n    try {\n      const resolution = this.documentExtraction.resolveBinding(DocumentExtraction_Applicability.DOCUMENT_TEXT_DEFAULT);\n      // `=== false` rather than `!…`: this compiler configuration narrows the literal-discriminant\n      // union on the explicit comparison but not on the truthiness form. Measured here too, not\n      // copied on faith — `!resolution.configured` compiles and then fails to narrow `reason`.\n      if (resolution.configured === false) {\n        this.logDebug('No document-extraction provider configured — using the local library', { fileId, reason: resolution.reason });\n        return undefined;\n      }\n      if (!this.documentExtraction.canServe(resolution.binding, mimeType, byteLength)) {\n        this.logDebug('Configured document-extraction provider cannot serve this document — using the local library', {\n          fileId,\n          mimeType,\n          byteLength,\n          maxDocumentBytes: resolution.binding.maxDocumentBytes,\n          providerRef: resolution.binding.providerRef,\n        });\n        return undefined;\n      }\n      return resolution.binding;\n    } catch (resolutionError) {\n      this.logger.warn('Document-extraction provider is misconfigured — falling back to the local library for this document', {\n        fileId,\n        reason: resolutionError instanceof Error ? resolutionError.message : String(resolutionError),\n      });\n      return undefined;\n    }\n  }\n\n  /**\n   * Extract text from a PDF using the LOCAL `pdfjs-dist` library.\n   *\n   * ## Why the import stays dynamic even though the dependency is installed\n   *\n   * `pdfjs-dist` is a large module graph that most backend processes never touch — only a request\n   * that actually carries a PDF needs it. A dynamic import keeps it out of every other process's\n   * startup cost. What changed is that the import can no longer FAIL for want of a dependency: the\n   * package is declared in this package's manifest, so a resolution failure now means a broken\n   * installation, which is a real fault and is correctly loud.\n   *\n   * The deep specifier targets the `legacy` build deliberately. It is the ESM build pdf.js ships for\n   * non-browser hosts; the default build assumes browser globals. `pdfjs-dist` publishes no\n   * `exports` map, so this deep path is a supported entrypoint rather than a boundary violation.\n   *\n   * ## When a PROVIDER serves this instead\n   *\n   * This local path reads the text layer a PDF already contains, and returns nothing useful for a\n   * SCANNED document (an image of a page carries no text layer). That is what the provider-backed\n   * OCR capability exists to handle, and {@link extractFromDocument} is the seam that chooses\n   * between them. With no `AI_DOCUMENT_EXTRACTION` provider configured, a scanned PDF still\n   * extracts to empty text — which the RAG ingestion path treats as \"contributes no chunks\"\n   * rather than as an error.\n   *\n   * It takes a BUFFER rather than a stream because the selection seam above may need the same\n   * bytes twice (provider first, this second), and a `Readable` is consumed once. `reason` names\n   * which rung of that seam routed here and is logged, so an operator reading the logs can tell\n   * \"no provider is configured\" from \"the provider was down\" without inspecting configuration.\n   */\n  /**\n   * The LOCAL-LIBRARY rung of {@link extractFromDocument}'s ladder, dispatched by MIME.\n   *\n   * Two libraries sit at this rung and they are not interchangeable: `pdfjs-dist` reads a PDF's\n   * text layer, and the ZIP+XML extractor reads OOXML / OpenDocument parts. Handing a document to\n   * the wrong one produces an unreadable-bytes classification that blames the FILE for a routing\n   * mistake — which is exactly the failure the provider-only era's refusal existed to avoid, so it\n   * must not reappear now that there are two locals instead of one.\n   *\n   * One dispatcher rather than a branch at each of the five call sites, because the five rungs are\n   * the same decision reached for five different reasons; branching at each would be five places to\n   * forget the office arm the day a sixth rung appears.\n   */\n  private async extractLocally(\n    buffer: Buffer,\n    fileId: string,\n    mimeType: string,\n    reason: string,\n  ): Promise<FileTextExtractionResult> {\n    if (LOCALLY_EXTRACTABLE_DOCUMENT_MIMES.has(mimeType.toLowerCase())) {\n      return await this.extractFromOfficeDocument(buffer, fileId, mimeType.toLowerCase(), reason);\n    }\n    return await this.extractFromPDF(buffer, fileId, reason);\n  }\n\n  /**\n   * OOXML / OpenDocument text, read locally (#434).\n   *\n   * `partCount` is reported as `pageCount` deliberately: for a `.pptx` it IS the slide count, and\n   * for the others it is the closest honest analogue of \"how many units of this document were\n   * read\". Inventing a page count by dividing characters would be a number an operator could\n   * reasonably act on and that means nothing.\n   *\n   * An {@link OfficeDocumentExtractionError} is rethrown UNTOUCHED so the caller's classifier sees\n   * it: every one of its failures is a statement about the document's own bytes — not a readable\n   * ZIP, larger than the decompression bound, or holding none of the parts its type keeps text in\n   * — which is precisely the unreadable-document bucket, and not the provider-unavailable one that\n   * must protect an existing corpus.\n   */\n  private async extractFromOfficeDocument(\n    buffer: Buffer,\n    fileId: string,\n    mimeType: string,\n    reason: string,\n  ): Promise<FileTextExtractionResult> {\n    const extracted = await extractOfficeDocumentText(buffer, mimeType);\n\n    this.logDebug('Extracted text from an Office/OpenDocument file with the local library', {\n      fileId,\n      mimeType,\n      reason,\n      partCount: extracted.partCount,\n      characterCount: extracted.text.length,\n    });\n\n    return {\n      text: extracted.text,\n      metadata: {\n        pageCount: extracted.partCount,\n        characterCount: extracted.text.length,\n        wordCount: extracted.text.split(/\\s+/).filter(Boolean).length,\n        extractionPath: 'local-library',\n      },\n    };\n  }\n\n  private async extractFromPDF(buffer: Buffer, fileId: string, reason: string): Promise<FileTextExtractionResult> {\n    const pdfjs = await import('pdfjs-dist/legacy/build/pdf.mjs');\n\n    const loadingTask = pdfjs.getDocument({\n      // A COPY, and it is mandatory rather than defensive. pdf.js transfers the backing\n      // ArrayBuffer to its worker port and DETACHES it. For a small allocation Node's `Buffer`\n      // is a window onto a SHARED 8 KB pool, so handing it `buffer` directly would detach a\n      // region other, unrelated buffers are still using. `new Uint8Array(buffer)` copies out of\n      // the pool into a private allocation. Costs 2x peak memory for the document; correctness\n      // is worth it.\n      data: new Uint8Array(buffer),\n      // Not the hardening it looks like — measured against the shipped build 2026-08-27, and\n      // recorded because the plausible reading is wrong in both directions.\n      //\n      // It does NOT stop a document reaching the host's font files: `useSystemFonts` only decides\n      // whether the worker builds a CSS `local()` substitution descriptor, which is string\n      // construction and opens nothing. And in Node it is ALREADY false by default.\n      //\n      // It is set explicitly for the opposite-facing reason: when `useSystemFonts` is true,\n      // `fetchStandardFontData` early-returns null, which would make the `standardFontDataUrl`\n      // below dead. Pinning it false is what keeps that parameter load-bearing.\n      //\n      // (A sibling `isEvalSupported: false` was written here first and removed: pdf.js 6 deleted\n      // that option outright — zero occurrences in its types AND its shipped build — having\n      // removed the eval-based font-compilation path. It type-errors, and would have advertised a\n      // decision the library silently ignored.)\n      useSystemFonts: false,\n      standardFontDataUrl: resolveStandardFontDataPath(),\n    });\n\n    try {\n      const pdfDocument = await loadingTask.promise;\n      const pageTexts: string[] = [];\n\n      for (let pageNumber = 1; pageNumber <= pdfDocument.numPages; pageNumber++) {\n        const page = await pdfDocument.getPage(pageNumber);\n        try {\n          const textContent = await page.getTextContent();\n          // Concatenate items HONOURING `hasEOL`, rather than joining every item with a space.\n          //\n          // pdf.js splits a visual line into several `TextItem`s at kerning and positioning\n          // boundaries, so a blanket `join(' ')` inserts spaces INSIDE words on ordinary documents\n          // and erases line structure entirely. That matters more here than it would in a viewer:\n          // this string is what gets chunked and embedded, so the damage lands in the retrieval\n          // corpus rather than on a screen where a human would notice it.\n          //\n          // `items` also mixes TextItem (which carries `str`) with TextMarkedContent (structural\n          // markers, which do not). Narrowing on the property rather than casting keeps a future\n          // pdf.js item kind from silently stringifying as `undefined`.\n          let pageText = '';\n          for (const item of textContent.items) {\n            if (!('str' in item)) continue;\n            pageText += item.str;\n            if (item.hasEOL) pageText += '\\n';\n          }\n          pageTexts.push(pageText);\n        } finally {\n          page.cleanup();\n        }\n      }\n\n      const text = pageTexts.join('\\n');\n\n      this.logDebug('Extracted text from PDF via the local pdfjs-dist path', {\n        fileId,\n        pageCount: pdfDocument.numPages,\n        characterCount: text.length,\n        reason,\n      });\n\n      return {\n        text,\n        metadata: {\n          pageCount: pdfDocument.numPages,\n          characterCount: text.length,\n          wordCount: text.split(/\\s+/).filter(Boolean).length,\n          // Which path served this extraction, and why. Carried in the RESULT rather than only in\n          // a log line so a caller (and a test) can assert the routing without scraping logs.\n          extractionPath: 'local-library',\n          extractionPathReason: reason,\n        },\n      };\n    } finally {\n      // `destroy()` lives on the LOADING TASK, not on the document proxy — `pdfDocument.destroy` is\n      // not a function in pdfjs-dist 6.x. Skipping this leaks the document's worker port.\n      //\n      // Swallowed deliberately: `destroy()` can itself throw, and a throw from a `finally` REPLACES\n      // the in-flight error. The caller classifies on that original error to decide whether a\n      // document is permanently unreadable or the storage layer is merely down, so losing it would\n      // corrupt exactly the decision this method exists to inform.\n      await loadingTask.destroy().catch(() => undefined);\n    }\n  }\n\n  /** Drain a download stream into a single buffer. Shared by every extraction path. */\n  private async collectStream(stream: Readable): Promise<Buffer> {\n    const chunks: Buffer[] = [];\n    for await (const chunk of stream) {\n      chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk));\n    }\n    return Buffer.concat(chunks);\n  }\n}\n"]}