@heripo/document-processor 0.1.4 → 0.1.5

package/dist/index.cjs

@@ -1268,7 +1268,8 @@ var TocValidationError = class extends TocExtractError {
 // src/extractors/toc-validator.ts
 var DEFAULT_OPTIONS = {
   totalPages: Infinity,
-  maxTitleLength: 200
+  maxTitleLength: 200,
+  maxFirstEntryPageRatio: 0.3
 };
 var TocValidator = class {
   options;
@@ -1289,6 +1290,7 @@ var TocValidator = class {
   validate(entries) {
     this.issues = [];
     this.validateEntries(entries, "", null, /* @__PURE__ */ new Set());
+    this.validateFirstEntryPagePosition(entries);
     const errorCount = this.issues.length;
     return {
       valid: errorCount === 0,
@@ -1305,8 +1307,12 @@ var TocValidator = class {
   validateOrThrow(entries) {
     const result = this.validate(entries);
     if (!result.valid) {
+      const details = result.issues.map(
+        (issue) => `  [${issue.code}] ${issue.message} (path: ${issue.path}, entry: "${issue.entry.title}" page ${issue.entry.pageNo})`
+      ).join("\n");
       throw new TocValidationError(
-        `TOC validation failed with ${result.errorCount} error(s)`,
+        `TOC validation failed with ${result.errorCount} error(s):
+${details}`,
         result
       );
     }
@@ -1421,6 +1427,33 @@ var TocValidator = class {
       });
     }
   }
+  /**
+   * V007: Validate first entry page position (completeness check)
+   *
+   * If the first level-1 entry starts too late in the document,
+   * earlier entries might be missing from the TOC.
+   */
+  validateFirstEntryPagePosition(entries) {
+    if (entries.length === 0) {
+      return;
+    }
+    if (!isFinite(this.options.totalPages)) {
+      return;
+    }
+    const firstEntry = entries[0];
+    const threshold = Math.max(
+      50,
+      Math.floor(this.options.totalPages * this.options.maxFirstEntryPageRatio)
+    );
+    if (firstEntry.pageNo > threshold) {
+      this.addIssue({
+        code: "V007",
+        message: `TOC may be incomplete - first entry starts at page ${firstEntry.pageNo}, expected within first ${threshold} pages. Earlier entries might be missing.`,
+        path: "[0]",
+        entry: firstEntry
+      });
+    }
+  }
   /**
    * Add issue to the list
    */
@@ -1639,22 +1672,42 @@ var TocFinder = class {
     return numberCount > 0 && numberCount / (num_rows - 1) > 0.5;
   }
   /**
-   * Expand TOC area to consecutive pages
+   * Expand TOC area to consecutive pages (both backward and forward)
    */
   expandToConsecutivePages(initial, doc) {
     const itemRefs = [...initial.itemRefs];
+    const seenRefs = new Set(itemRefs);
+    let startPage = initial.startPage;
     let endPage = initial.endPage;
+    for (let pageNo = initial.startPage - 1; pageNo >= 1; pageNo--) {
+      const continuationItems = this.findContinuationOnPage(doc, pageNo);
+      if (continuationItems.length === 0) {
+        break;
+      }
+      const newItems = continuationItems.filter((ref) => !seenRefs.has(ref));
+      for (const ref of newItems) {
+        seenRefs.add(ref);
+      }
+      itemRefs.unshift(...newItems);
+      startPage = pageNo;
+      this.logger.info(`[TocFinder] Expanded TOC backward to page ${pageNo}`);
+    }
     for (let pageNo = initial.endPage + 1; pageNo <= this.maxSearchPages; pageNo++) {
       const continuationItems = this.findContinuationOnPage(doc, pageNo);
       if (continuationItems.length === 0) {
         break;
       }
-      itemRefs.push(...continuationItems);
+      const newItems = continuationItems.filter((ref) => !seenRefs.has(ref));
+      for (const ref of newItems) {
+        seenRefs.add(ref);
+      }
+      itemRefs.push(...newItems);
       endPage = pageNo;
+      this.logger.info(`[TocFinder] Expanded TOC forward to page ${pageNo}`);
     }
     return {
       itemRefs,
-      startPage: initial.startPage,
+      startPage,
       endPage
     };
   }
@@ -1872,12 +1925,22 @@ var TextLLMComponent = class extends BaseLLMComponent {
 };
 
 // src/extractors/toc-extractor.ts
+var MAX_VALIDATION_RETRIES = 3;
+var VALIDATION_CODE_DESCRIPTIONS = {
+  V001: "Page numbers must be in non-decreasing order within the same level. A decrease usually means a hierarchy or page number error.",
+  V002: "Page number is out of valid range (must be >= 1 and <= total pages).",
+  V003: "Title is empty or contains only whitespace.",
+  V004: "Title exceeds the maximum allowed length.",
+  V005: "Child page number is before parent page number. Children must start on or after the parent page.",
+  V006: "Duplicate entry detected (same title and page number).",
+  V007: "First TOC entry starts too late in the document. Earlier entries may be missing."
+};
 var TocEntrySchema = import_zod.z.lazy(
   () => import_zod.z.object({
     title: import_zod.z.string().describe("Chapter or section title"),
     level: import_zod.z.number().int().min(1).describe("Hierarchy depth (1 = top level)"),
     pageNo: import_zod.z.number().int().min(1).describe("Starting page number"),
-    children: import_zod.z.array(TocEntrySchema).optional().describe("Child sections")
+    children: import_zod.z.array(TocEntrySchema).describe("Child sections (use empty array [] if none)")
   })
 );
 var TocResponseSchema = import_zod.z.object({
@@ -1900,12 +1963,15 @@ var TocExtractor = class extends TextLLMComponent {
   /**
    * Extract TOC structure from Markdown
    *
+   * When validation fails, retries with correction feedback up to MAX_VALIDATION_RETRIES times.
+   *
    * @param markdown - Markdown representation of TOC area
-   * @returns Object with entries array and token usage information
+   * @param validationOverrides - Optional overrides for validation options (merged with constructor options)
+   * @returns Object with entries array and token usage array (initial extraction + any corrections)
    * @throws {TocParseError} When LLM fails to parse structure
-   * @throws {TocValidationError} When validation fails
+   * @throws {TocValidationError} When validation fails after all retries
    */
-  async extract(markdown) {
+  async extract(markdown, validationOverrides) {
     this.log("info", `Starting TOC extraction (${markdown.length} chars)`);
     if (!markdown.trim()) {
       this.log("error", "Cannot extract TOC from empty markdown content");
@@ -1920,18 +1986,52 @@ var TocExtractor = class extends TextLLMComponent {
         this.buildUserPrompt(markdown),
         "extraction"
       );
-      const entries = this.normalizeEntries(result.output.entries);
+      const usages = [result.usage];
+      let entries = this.normalizeEntries(result.output.entries);
       if (!this.skipValidation) {
-        this.validateEntries(entries);
+        let validationError = this.tryValidateEntries(
+          entries,
+          validationOverrides
+        );
+        for (let attempt = 1; attempt <= MAX_VALIDATION_RETRIES && validationError !== null; attempt++) {
+          this.log(
+            "warn",
+            `Validation failed (attempt ${attempt}/${MAX_VALIDATION_RETRIES}), retrying with correction feedback`
+          );
+          const correctionPrompt = this.buildCorrectionPrompt(
+            markdown,
+            entries,
+            validationError.validationResult.issues
+          );
+          const correctionResult = await this.callTextLLM(
+            TocResponseSchema,
+            this.buildSystemPrompt(),
+            correctionPrompt,
+            `correction-${attempt}`
+          );
+          usages.push(correctionResult.usage);
+          entries = this.normalizeEntries(correctionResult.output.entries);
+          validationError = this.tryValidateEntries(
+            entries,
+            validationOverrides
+          );
+        }
+        if (validationError !== null) {
+          this.log(
+            "error",
+            `Validation failed after ${MAX_VALIDATION_RETRIES} retries:
+${validationError.getSummary()}`
+          );
+          throw validationError;
+        }
       }
       this.log(
         "info",
-        `Extraction completed: ${entries.length} top-level entries`
+        `Extraction completed: ${entries.length} top-level entries (${usages.length} LLM call(s))`
       );
-      return { entries, usage: result.usage };
+      return { entries, usages };
     } catch (error) {
       if (error instanceof TocValidationError) {
-        this.log("error", `Validation failed: ${error.message}`);
         throw error;
       }
       const message = error instanceof Error ? error.message : String(error);
@@ -1942,16 +2042,69 @@ var TocExtractor = class extends TextLLMComponent {
     }
   }
   /**
-   * Validate extracted entries
+   * Validate extracted entries and return error or null
    *
-   * @throws {TocValidationError} When validation fails
+   * Unlike validateOrThrow, this returns the error instead of throwing,
+   * allowing the retry loop to handle it.
+   *
+   * @returns TocValidationError if validation fails, null if valid
    */
-  validateEntries(entries) {
+  tryValidateEntries(entries, overrides) {
     if (entries.length === 0) {
-      return;
+      return null;
     }
-    const validator = new TocValidator(this.validationOptions);
-    validator.validateOrThrow(entries);
+    const options = { ...this.validationOptions, ...overrides };
+    const validator = new TocValidator(options);
+    const result = validator.validate(entries);
+    if (!result.valid) {
+      const details = result.issues.map(
+        (issue) => `  [${issue.code}] ${issue.message} (path: ${issue.path}, entry: "${issue.entry.title}" page ${issue.entry.pageNo})`
+      ).join("\n");
+      return new TocValidationError(
+        `TOC validation failed with ${result.errorCount} error(s):
+${details}`,
+        result
+      );
+    }
+    return null;
+  }
+  /**
+   * Build correction prompt with validation error feedback
+   *
+   * Includes the original markdown, previous extraction result,
+   * validation errors, and guidance for fixing common mistakes.
+   */
+  buildCorrectionPrompt(markdown, previousEntries, issues) {
+    const errorLines = issues.map((issue) => {
+      const desc = VALIDATION_CODE_DESCRIPTIONS[issue.code] ?? "Unknown validation error.";
+      return `- [${issue.code}] ${issue.message}
+  Path: ${issue.path}
+  Entry: "${issue.entry.title}" (page ${issue.entry.pageNo})
+  Rule: ${desc}`;
+    });
+    return `Your previous TOC extraction had validation errors. Please fix them and re-extract.
+
+## Validation Errors
+
+${errorLines.join("\n\n")}
+
+## Common Mistakes to Avoid
+
+1. **Hierarchy confusion**: Entries with the same numbering prefix (e.g., "4)") can belong to different hierarchy levels depending on context. Use indentation and surrounding entries to determine the correct parent-child relationship.
+2. **Page number misread**: Carefully distinguish Roman numerals (VI=6) from Arabic numerals. "VI. \uACE0\uCC30" at page 277 is NOT "V. \uACE0\uCC30" at page 27.
+3. **Page order**: Within the same parent, sibling entries must have non-decreasing page numbers. If a page number decreases, the entry likely belongs to a different hierarchy level.
+
+## Original Markdown
+
+${markdown}
+
+## Your Previous Extraction (with errors)
+
+${JSON.stringify(previousEntries, null, 2)}
+
+## Instructions
+
+Re-extract the TOC structure from the original markdown above. Fix all validation errors listed above. Return the corrected entries.`;
   }
   /**
    * Build system prompt for TOC extraction
@@ -1969,11 +2122,12 @@ var TocExtractor = class extends TextLLMComponent {
    - Level 3: Subsections (e.g., "1.1.1", "a.", "(1)")
    - Use indentation and numbering patterns to infer level
 
-3. **Page Number**: Extract the page number from each entry. Convert Roman numerals to Arabic numerals if present (e.g., "iv" \u2192 4).
+3. **Page Number**: Extract the page number from each entry. Use only Arabic numerals for page numbers.
 
 4. **Children**: Nest child entries under parent entries based on their hierarchy level.
 
-5. **IMPORTANT - Extract Main TOC Only**: Only extract the main document table of contents. EXCLUDE the following supplementary indices:
+5. **IMPORTANT - Extract Main TOC Only**: Only extract the main document table of contents. EXCLUDE the following:
+   - **Front matter with Roman numeral pages**: Entries whose page numbers are Roman numerals (i, ii, xxi, etc.) such as \uC77C\uB7EC\uB450\uAE30, \uBC1C\uAC04\uC0AC, \uC11C\uBB38, \uBC94\uB840, Preface, Foreword, Editorial Notes. These use a separate page numbering system and are not part of the main content.
    - Photo/image indices (\uC0AC\uC9C4 \uBAA9\uCC28, \uC0AC\uC9C4\uBAA9\uCC28, \uD654\uBCF4 \uBAA9\uCC28, Photo Index, List of Photos, List of Figures)
    - Drawing/diagram indices (\uB3C4\uBA74 \uBAA9\uCC28, \uB3C4\uBA74\uBAA9\uCC28, \uC0BD\uB3C4 \uBAA9\uCC28, Drawing Index, List of Drawings)
    - Table indices (\uD45C \uBAA9\uCC28, \uD45C\uBAA9\uCC28, Table Index, List of Tables)
@@ -2000,11 +2154,11 @@ Output:
       "level": 1,
       "pageNo": 1,
       "children": [
-        { "title": "1. \uC5F0\uAD6C \uBC30\uACBD", "level": 2, "pageNo": 3 },
-        { "title": "2. \uC5F0\uAD6C \uBAA9\uC801", "level": 2, "pageNo": 5 }
+        { "title": "1. \uC5F0\uAD6C \uBC30\uACBD", "level": 2, "pageNo": 3, "children": [] },
+        { "title": "2. \uC5F0\uAD6C \uBAA9\uC801", "level": 2, "pageNo": 5, "children": [] }
       ]
     },
-    { "title": "\uC81C2\uC7A5 \uBC29\uBC95\uB860", "level": 1, "pageNo": 10 }
+    { "title": "\uC81C2\uC7A5 \uBC29\uBC95\uB860", "level": 1, "pageNo": 10, "children": [] }
   ]
 }`;
   }
@@ -2647,7 +2801,7 @@ var PagePattern = /* @__PURE__ */ ((PagePattern2) => {
 var PageRangeParser = class extends VisionLLMComponent {
   // Configuration constants
   SAMPLE_SIZE = 3;
-  MAX_PATTERN_RETRIES = 6;
+  MAX_PATTERN_RETRIES = 19;
   SIZE_TOLERANCE = 5;
   constructor(logger, model, outputPath, maxRetries = 3, fallbackModel, aggregator, abortSignal) {
     super(
@@ -4047,8 +4201,8 @@ var DocumentProcessor = class {
     }
     if (!markdown) {
       this.logger.info("[DocumentProcessor] Using vision fallback for TOC");
-      const totalPages = Object.keys(doclingDoc.pages).length;
-      markdown = await this.visionTocExtractor.extract(totalPages);
+      const totalPages2 = Object.keys(doclingDoc.pages).length;
+      markdown = await this.visionTocExtractor.extract(totalPages2);
       if (!markdown) {
         const reason = "Both rule-based search and vision fallback failed to locate TOC";
         this.logger.error(
@@ -4062,8 +4216,13 @@ var DocumentProcessor = class {
         `[DocumentProcessor] Vision extracted TOC markdown (${markdown.length} chars)`
       );
     }
-    const tocResult = await this.tocExtractor.extract(markdown);
-    this.usageAggregator.track(tocResult.usage);
+    const totalPages = Object.keys(doclingDoc.pages).length;
+    const tocResult = await this.tocExtractor.extract(markdown, {
+      totalPages
+    });
+    for (const usage of tocResult.usages) {
+      this.usageAggregator.track(usage);
+    }
     if (tocResult.entries.length === 0) {
       const reason = "TOC area was detected but LLM could not extract any structured entries";
       this.logger.error(`[DocumentProcessor] TOC extraction failed: ${reason}`);