docrev 0.9.13 → 0.9.14

package/lib/word-extraction.ts

@@ -1,806 +1,806 @@
-/**
- * Word document data extraction - raw extraction from .docx files
- */
-
-import * as fs from 'fs';
-import * as path from 'path';
-import { exec } from 'child_process';
-import { promisify } from 'util';
-
-const execAsync = promisify(exec);
-
-// ============================================
-// Type Definitions
-// ============================================
-
-export interface WordComment {
-  id: string;
-  author: string;
-  date: string;
-  text: string;
-  /**
-   * Parent comment id when this is a reply in a Word comment thread.
-   * Resolved from `commentsExtended.xml`'s `w15:paraIdParent` field.
-   * `undefined` for top-level comments.
-   */
-  parentId?: string;
-}
-
-export interface TextNode {
-  xmlStart: number;
-  xmlEnd: number;
-  textStart: number;
-  textEnd: number;
-  text: string;
-}
-
-export interface CommentAnchorData {
-  anchor: string;
-  before: string;
-  after: string;
-  docPosition: number;
-  docLength: number;
-  isEmpty: boolean;
-}
-
-export interface CommentAnchorsResult {
-  anchors: Map<string, CommentAnchorData>;
-  fullDocText: string;
-}
-
-export interface DocxHeading {
-  /** Heading style name from `<w:pStyle>`, e.g. "Heading1" */
-  style: string;
-  /** Heading depth: 1, 2, 3, ... (parsed from style name; 0 if unknown) */
-  level: number;
-  /** Concatenated text content of the heading paragraph */
-  text: string;
-  /** Position in fullDocText (same coordinate system as CommentAnchorData.docPosition) */
-  docPosition: number;
-}
-
-export interface WordTable {
-  markdown: string;
-  rowCount: number;
-  colCount: number;
-}
-
-export interface ParsedRow {
-  cells: string[];
-  colSpans: number[];
-}
-
-export interface ExtractFromWordOptions {
-  mediaDir?: string;
-  skipMediaExtraction?: boolean;
-}
-
-export interface ExtractMessage {
-  type: 'info' | 'warning';
-  message: string;
-}
-
-export interface ExtractFromWordResult {
-  text: string;
-  comments: WordComment[];
-  anchors: Map<string, CommentAnchorData>;
-  messages: ExtractMessage[];
-  extractedMedia: string[];
-  tables: WordTable[];
-  hasTrackChanges: boolean;
-  trackChangeStats: { insertions: number; deletions: number };
-}
-
-// ============================================
-// Functions
-// ============================================
-
-/**
- * Extract comments directly from Word docx comments.xml
- */
-export async function extractWordComments(docxPath: string): Promise<WordComment[]> {
-  const AdmZip = (await import('adm-zip')).default;
-  const { parseStringPromise } = await import('xml2js');
-
-  const comments: WordComment[] = [];
-
-  // Validate file exists
-  if (!fs.existsSync(docxPath)) {
-    throw new Error(`File not found: ${docxPath}`);
-  }
-
-  try {
-    let zip;
-    try {
-      zip = new AdmZip(docxPath);
-    } catch (err: any) {
-      throw new Error(`Invalid Word document (not a valid .docx file): ${err.message}`);
-    }
-
-    const commentsEntry = zip.getEntry('word/comments.xml');
-
-    if (!commentsEntry) {
-      return comments;
-    }
-
-    let commentsXml;
-    try {
-      commentsXml = commentsEntry.getData().toString('utf8');
-    } catch (err: any) {
-      throw new Error(`Failed to read comments from document: ${err.message}`);
-    }
-
-    const parsed = await parseStringPromise(commentsXml, { explicitArray: false });
-
-    const commentsRoot = parsed['w:comments'];
-    if (!commentsRoot || !commentsRoot['w:comment']) {
-      return comments;
-    }
-
-    // Ensure it's an array
-    const commentNodes = Array.isArray(commentsRoot['w:comment'])
-      ? commentsRoot['w:comment']
-      : [commentsRoot['w:comment']];
-
-    // Map every paraId that lives inside a comment back to that comment's id.
-    // Word's commentsExtended.xml expresses threading via w15:paraIdParent,
-    // which references the parent's first <w:p>. Replies use a secondary
-    // (often-empty) <w:p>, so each comment may contribute multiple paraIds.
-    const paraIdToCommentId = new Map<string, string>();
-
-    for (const comment of commentNodes) {
-      const id = comment.$?.['w:id'] || '';
-      const author = comment.$?.['w:author'] || 'Unknown';
-      const date = comment.$?.['w:date'] || '';
-
-      // Extract text from nested w:p/w:r/w:t elements and record paraIds.
-      let text = '';
-      const extractText = (node: any): void => {
-        if (!node) return;
-        if (typeof node === 'string') {
-          text += node;
-          return;
-        }
-        if (node['w:t']) {
-          const t = node['w:t'];
-          text += typeof t === 'string' ? t : (t._ || t);
-        }
-        if (node['w:r']) {
-          const runs = Array.isArray(node['w:r']) ? node['w:r'] : [node['w:r']];
-          runs.forEach(extractText);
-        }
-        if (node['w:p']) {
-          const paras = Array.isArray(node['w:p']) ? node['w:p'] : [node['w:p']];
-          for (const para of paras) {
-            const paraId = para?.$?.['w14:paraId'];
-            if (paraId && id) paraIdToCommentId.set(paraId, id);
-            extractText(para);
-          }
-        }
-      };
-      extractText(comment);
-
-      comments.push({ id, author, date: date.slice(0, 10), text: text.trim() });
-    }
-
-    // Resolve parent links from commentsExtended.xml. Missing entry just
-    // means the docx has no threading metadata (e.g. legacy/non-Word source).
-    const extendedEntry = zip.getEntry('word/commentsExtended.xml');
-    if (extendedEntry && paraIdToCommentId.size > 0) {
-      let extendedXml = '';
-      try {
-        extendedXml = extendedEntry.getData().toString('utf8');
-      } catch {
-        // Unreadable threading metadata is non-fatal; skip parent linking.
-      }
-      if (extendedXml) {
-        const parentByCommentId = new Map<string, string>();
-        const exPattern = /<w15:commentEx\b([^>]*?)\/>/g;
-        let m: RegExpExecArray | null;
-        while ((m = exPattern.exec(extendedXml)) !== null) {
-          const attrs = m[1] ?? '';
-          const paraIdMatch = attrs.match(/w15:paraId="([^"]+)"/);
-          const parentMatch = attrs.match(/w15:paraIdParent="([^"]+)"/);
-          if (!paraIdMatch || !parentMatch) continue;
-          const childCommentId = paraIdToCommentId.get(paraIdMatch[1]);
-          const parentCommentId = paraIdToCommentId.get(parentMatch[1]);
-          if (childCommentId && parentCommentId && childCommentId !== parentCommentId) {
-            parentByCommentId.set(childCommentId, parentCommentId);
-          }
-        }
-        for (const c of comments) {
-          const parent = parentByCommentId.get(c.id);
-          if (parent) c.parentId = parent;
-        }
-      }
-    }
-  } catch (err: any) {
-    // Re-throw with more context if it's already an Error we created
-    if (err.message.includes('Invalid Word document') || err.message.includes('File not found')) {
-      throw err;
-    }
-    throw new Error(`Error extracting comments from ${path.basename(docxPath)}: ${err.message}`);
-  }
-
-  return comments;
-}
-
-/**
- * Extract comment anchor texts from document.xml with surrounding context
- * Returns map of comment ID -> {anchor, before, after, docPosition, isEmpty} for better matching
- * Also returns fullDocText for section boundary matching
- */
-export async function extractCommentAnchors(docxPath: string): Promise<CommentAnchorsResult> {
-  const AdmZip = (await import('adm-zip')).default;
-  const anchors = new Map<string, CommentAnchorData>();
-  let fullDocText = '';
-
-  try {
-    const zip = new AdmZip(docxPath);
-    const docEntry = zip.getEntry('word/document.xml');
-
-    if (!docEntry) {
-      return { anchors, fullDocText };
-    }
-
-    const docXml = docEntry.getData().toString('utf8');
-
-    // ========================================
-    // STEP 1: Build text position mapping
-    // ========================================
-    const textNodePattern = /<w:t[^>]*>([^<]*)<\/w:t>/g;
-    const textNodes: TextNode[] = [];
-    let textPosition = 0;
-    let nodeMatch;
-
-    while ((nodeMatch = textNodePattern.exec(docXml)) !== null) {
-      const rawText = nodeMatch[1] ?? '';
-      const decodedText = decodeXmlEntities(rawText);
-      textNodes.push({
-        xmlStart: nodeMatch.index,
-        xmlEnd: nodeMatch.index + nodeMatch[0].length,
-        textStart: textPosition,
-        textEnd: textPosition + decodedText.length,
-        text: decodedText
-      });
-      textPosition += decodedText.length;
-    }
-
-    fullDocText = textNodes.map(n => n.text).join('');
-
-    // Helper: convert XML position to text position
-    function xmlPosToTextPos(xmlPos: number): number {
-      for (let i = 0; i < textNodes.length; i++) {
-        const node = textNodes[i];
-        if (!node) continue;
-        if (xmlPos >= node.xmlStart && xmlPos < node.xmlEnd) {
-          return node.textStart;
-        }
-        if (xmlPos < node.xmlStart) {
-          return node.textStart;
-        }
-      }
-      const lastNode = textNodes[textNodes.length - 1];
-      return lastNode ? lastNode.textEnd : 0;
-    }
-
-    // Helper: extract context before a position
-    function getContextBefore(position: number, maxLength: number = 150): string {
-      const beforeText = fullDocText.slice(Math.max(0, position - maxLength), position);
-      const sentenceStart = beforeText.search(/[.!?]\s+[A-Z][^.!?]*$/);
-      return sentenceStart >= 0
-        ? beforeText.slice(sentenceStart + 2).trim()
-        : beforeText.slice(-80).trim();
-    }
-
-    // Helper: extract context after a position
-    function getContextAfter(position: number, maxLength: number = 150): string {
-      const afterText = fullDocText.slice(position, position + maxLength);
-      const sentenceEnd = afterText.search(/[.!?]\s/);
-      return sentenceEnd >= 0
-        ? afterText.slice(0, sentenceEnd + 1).trim()
-        : afterText.slice(0, 80).trim();
-    }
-
-    // ========================================
-    // STEP 2: Collect all start/end markers separately
-    // ========================================
-    const startPattern = /<w:commentRangeStart[^>]*w:id="(\d+)"[^>]*\/?>/g;
-    const endPattern = /<w:commentRangeEnd[^>]*w:id="(\d+)"[^>]*\/?>/g;
-
-    const starts = new Map<string, number>();  // id -> position after start tag
-    const ends = new Map<string, number>();    // id -> position before end tag
-
-    let match;
-    while ((match = startPattern.exec(docXml)) !== null) {
-      const id = match[1];
-      if (!starts.has(id)) {
-        starts.set(id, match.index + match[0].length);
-      }
-    }
-
-    while ((match = endPattern.exec(docXml)) !== null) {
-      const id = match[1];
-      if (!ends.has(id)) {
-        ends.set(id, match.index);
-      }
-    }
-
-    // ========================================
-    // STEP 3: Process each comment range by ID
-    // ========================================
-    for (const [id, startXmlPos] of starts) {
-      const endXmlPos = ends.get(id);
-
-      // Missing end marker - skip with warning
-      if (endXmlPos === undefined) {
-        console.warn(`Comment ${id}: missing end marker`);
-        continue;
-      }
-
-      // Calculate text position
-      const docPosition = xmlPosToTextPos(startXmlPos);
-
-      // Handle empty or inverted ranges
-      if (endXmlPos <= startXmlPos) {
-        anchors.set(id, {
-          anchor: '',
-          before: getContextBefore(docPosition),
-          after: getContextAfter(docPosition),
-          docPosition,
-          docLength: fullDocText.length,
-          isEmpty: true
-        });
-        continue;
-      }
-
-      // Extract XML segment between markers
-      const segment = docXml.slice(startXmlPos, endXmlPos);
-
-      // Extract text from w:t (regular) AND w:delText (deleted text in track changes)
-      const textInRangePattern = /<w:t[^>]*>([^<]*)<\/w:t>|<w:delText[^>]*>([^<]*)<\/w:delText>/g;
-      let anchorText = '';
-      let tm;
-      while ((tm = textInRangePattern.exec(segment)) !== null) {
-        anchorText += tm[1] || tm[2] || '';
-      }
-      anchorText = decodeXmlEntities(anchorText);
-
-      // Get context
-      const anchorLength = anchorText.length;
-      const before = getContextBefore(docPosition);
-      const after = getContextAfter(docPosition + anchorLength);
-
-      // ALWAYS add entry (even if anchor is empty)
-      anchors.set(id, {
-        anchor: anchorText.trim(),
-        before,
-        after,
-        docPosition,
-        docLength: fullDocText.length,
-        isEmpty: !anchorText.trim()
-      });
-    }
-  } catch (err: any) {
-    console.error('Error extracting comment anchors:', err.message);
-    return { anchors, fullDocText: '' };
-  }
-
-  return { anchors, fullDocText };
-}
-
-/**
- * Extract heading paragraphs from a docx, with their text positions in the
- * same coordinate system as `extractCommentAnchors`'s `fullDocText` and
- * `CommentAnchorData.docPosition`.
- *
- * Headings are paragraphs whose `<w:pStyle>` is a Heading style. Reading
- * styles directly is more reliable than keyword-matching the concatenated
- * body text — there, paragraph boundaries are gone, so the literal string
- * "Methods" can appear inside prose ("results across countries") and the
- * structured-abstract label "Methods:" loses its colon when text runs are
- * concatenated.
- */
-export async function extractHeadings(docxPath: string): Promise<DocxHeading[]> {
-  const AdmZip = (await import('adm-zip')).default;
-
-  if (!fs.existsSync(docxPath)) {
-    throw new Error(`File not found: ${docxPath}`);
-  }
-
-  const zip = new AdmZip(docxPath);
-  const docEntry = zip.getEntry('word/document.xml');
-  if (!docEntry) return [];
-  const xml = docEntry.getData().toString('utf8');
-
-  // Build the same xml-pos → text-pos mapping that extractCommentAnchors does
-  const textNodePattern = /<w:t[^>]*>([^<]*)<\/w:t>/g;
-  const nodes: Array<{ xmlStart: number; xmlEnd: number; textStart: number; textEnd: number }> = [];
-  let textPos = 0;
-  let m;
-  while ((m = textNodePattern.exec(xml)) !== null) {
-    const decoded = decodeXmlEntities(m[1] ?? '');
-    nodes.push({
-      xmlStart: m.index,
-      xmlEnd: m.index + m[0].length,
-      textStart: textPos,
-      textEnd: textPos + decoded.length,
-    });
-    textPos += decoded.length;
-  }
-
-  function xmlToTextPos(xmlPos: number): number {
-    for (const n of nodes) {
-      if (xmlPos >= n.xmlStart && xmlPos < n.xmlEnd) return n.textStart;
-      if (xmlPos < n.xmlStart) return n.textStart;
-    }
-    return nodes.length ? nodes[nodes.length - 1].textEnd : 0;
-  }
-
-  const headings: DocxHeading[] = [];
-  const paraPattern = /<w:p\b[^>]*>([\s\S]*?)<\/w:p>/g;
-  let pm;
-  while ((pm = paraPattern.exec(xml)) !== null) {
-    const inner = pm[1];
-    const styleMatch = inner.match(/<w:pStyle[^>]*w:val="([^"]+)"/);
-    if (!styleMatch) continue;
-    const style = styleMatch[1];
-    if (!/heading/i.test(style)) continue;
-
-    // Concatenate text runs; include w:delText so a heading inside a tracked
-    // deletion is still surfaced (verifying anchors against an original draft)
-    const textInRange = /<w:t[^>]*>([^<]*)<\/w:t>|<w:delText[^>]*>([^<]*)<\/w:delText>/g;
-    let txt = '';
-    let tm;
-    while ((tm = textInRange.exec(inner)) !== null) {
-      txt += decodeXmlEntities(tm[1] || tm[2] || '');
-    }
-    const trimmed = txt.trim();
-    if (!trimmed) continue;
-
-    const levelMatch = style.match(/(\d+)/);
-    const level = levelMatch ? parseInt(levelMatch[1], 10) : 0;
-    headings.push({
-      style,
-      level,
-      text: trimmed,
-      docPosition: xmlToTextPos(pm.index),
-    });
-  }
-
-  return headings;
-}
-
-/**
- * Decode XML entities in text
- */
-function decodeXmlEntities(text: string): string {
-  return text
-    .replace(/&amp;/g, '&')
-    .replace(/&lt;/g, '<')
-    .replace(/&gt;/g, '>')
-    .replace(/&quot;/g, '"')
-    .replace(/&apos;/g, "'")
-    .replace(/&#(\d+);/g, (_, code) => String.fromCharCode(parseInt(code, 10)))
-    .replace(/&#x([0-9a-fA-F]+);/g, (_, code) => String.fromCharCode(parseInt(code, 16)));
-}
-
-/**
- * Extract text content from a Word XML cell
- */
-function extractCellText(cellXml: string): string {
-  const parts: string[] = [];
-
-  // Check for OMML math - replace with [math] placeholder
-  if (cellXml.includes('<m:oMath')) {
-    // Try to extract the text representation of math
-    const mathTextMatches = cellXml.match(/<m:t>([^<]*)<\/m:t>/g) || [];
-    if (mathTextMatches.length > 0) {
-      const mathText = mathTextMatches.map((t) => t.replace(/<[^>]+>/g, '')).join('');
-      parts.push(mathText);
-    } else {
-      parts.push('[math]');
-    }
-  }
-
-  // Extract regular text from w:t elements
-  const textMatches = cellXml.match(/<w:t[^>]*>([^<]*)<\/w:t>/g) || [];
-  for (const match of textMatches) {
-    const text = match.replace(/<[^>]+>/g, '');
-    if (text) {
-      parts.push(text);
-    }
-  }
-
-  let result = parts.join('').trim();
-  result = decodeXmlEntities(result);
-
-  // Escape pipe characters in cell content (would break table)
-  result = result.replace(/\|/g, '\\|');
-
-  return result;
-}
-
-/**
- * Parse a table row, handling merged cells (gridSpan)
- */
-function parseTableRow(rowXml: string, expectedCols: number): ParsedRow {
-  // Match cells - handle both <w:tc> and <w:tc ...>
-  const cellMatches = rowXml.match(/<w:tc(?:\s[^>]*)?>[\s\S]*?<\/w:tc>/g) || [];
-  const cells: string[] = [];
-  const colSpans: number[] = [];
-
-  for (const cellXml of cellMatches) {
-    // Check for horizontal merge (gridSpan)
-    const gridSpanMatch = cellXml.match(/<w:gridSpan\s+w:val="(\d+)"/);
-    const span = gridSpanMatch ? parseInt(gridSpanMatch[1], 10) : 1;
-
-    // Check for vertical merge continuation (vMerge without restart)
-    // If vMerge is present without w:val="restart", it's a continuation - use empty
-    const vMergeMatch = cellXml.match(/<w:vMerge(?:\s+w:val="([^"]+)")?/);
-    const isVMergeContinuation = vMergeMatch && vMergeMatch[1] !== 'restart';
-
-    const cellText = isVMergeContinuation ? '' : extractCellText(cellXml);
-
-    // Add the cell content
-    cells.push(cellText);
-    colSpans.push(span);
-
-    // For gridSpan > 1, add empty cells to maintain column alignment
-    for (let i = 1; i < span; i++) {
-      cells.push('');
-      colSpans.push(0); // 0 indicates this is a spanned cell
-    }
-  }
-
-  return { cells, colSpans };
-}
-
-/**
- * Determine table grid column count from table XML
- */
-function getTableGridCols(tableXml: string): number {
-  // Try to get from tblGrid
-  const gridColMatches = tableXml.match(/<w:gridCol/g) || [];
-  if (gridColMatches.length > 0) {
-    return gridColMatches.length;
-  }
-
-  // Fallback: count max cells in any row
-  const rowMatches = tableXml.match(/<w:tr[\s\S]*?<\/w:tr>/g) || [];
-  let maxCols = 0;
-  for (const rowXml of rowMatches) {
-    const { cells } = parseTableRow(rowXml, 0);
-    maxCols = Math.max(maxCols, cells.length);
-  }
-  return maxCols;
-}
-
-/**
- * Extract tables directly from Word document XML and convert to markdown pipe tables
- */
-export async function extractWordTables(docxPath: string): Promise<WordTable[]> {
-  const AdmZip = (await import('adm-zip')).default;
-  const tables: WordTable[] = [];
-
-  try {
-    const zip = new AdmZip(docxPath);
-    const docEntry = zip.getEntry('word/document.xml');
-
-    if (!docEntry) {
-      return tables;
-    }
-
-    const xml = docEntry.getData().toString('utf8');
-
-    // Find all table elements
-    const tableMatches = xml.match(/<w:tbl>[\s\S]*?<\/w:tbl>/g) || [];
-
-    for (const tableXml of tableMatches) {
-      // Determine expected column count from grid
-      const expectedCols = getTableGridCols(tableXml);
-
-      // Extract rows
-      const rowMatches = tableXml.match(/<w:tr[\s\S]*?<\/w:tr>/g) || [];
-      const rows: string[][] = [];
-
-      for (const rowXml of rowMatches) {
-        const { cells } = parseTableRow(rowXml, expectedCols);
-        if (cells.length > 0) {
-          rows.push(cells);
-        }
-      }
-
-      if (rows.length > 0) {
-        // Convert to markdown pipe table
-        const markdown = convertRowsToMarkdownTable(rows);
-        tables.push({ markdown, rowCount: rows.length, colCount: expectedCols || rows[0]?.length || 0 });
-      }
-    }
-  } catch (err: any) {
-    console.error('Error extracting tables from Word:', err.message);
-  }
-
-  return tables;
-}
-
-/**
- * Convert array of rows (each row is array of cell strings) to markdown pipe table
- */
-function convertRowsToMarkdownTable(rows: string[][]): string {
-  if (rows.length === 0) return '';
-
-  // Normalize column count (use max across all rows)
-  const colCount = Math.max(...rows.map((r) => r.length));
-
-  // Pad rows to have consistent column count
-  const normalizedRows = rows.map((row) => {
-    while (row.length < colCount) {
-      row.push('');
-    }
-    return row;
-  });
-
-  // Build markdown table
-  const lines: string[] = [];
-
-  // Header row
-  const header = normalizedRows[0];
-  lines.push('| ' + header.join(' | ') + ' |');
-
-  // Separator row
-  lines.push('|' + header.map(() => '---').join('|') + '|');
-
-  // Data rows
-  for (let i = 1; i < normalizedRows.length; i++) {
-    lines.push('| ' + normalizedRows[i].join(' | ') + ' |');
-  }
-
-  return lines.join('\n');
-}
-
-/**
- * Extract text from Word document using pandoc with track changes preserved
- */
-export async function extractFromWord(
-  docxPath: string,
-  options: ExtractFromWordOptions = {}
-): Promise<ExtractFromWordResult> {
-  let text: string;
-  let messages: ExtractMessage[] = [];
-  let extractedMedia: string[] = [];
-  let hasTrackChanges = false;
-  let trackChangeStats = { insertions: 0, deletions: 0 };
-
-  // Determine media extraction directory
-  const docxDir = path.dirname(docxPath);
-  const mediaDir = options.mediaDir || path.join(docxDir, 'media');
-
-  // Skip media extraction if figures already exist (e.g., when re-importing with existing source)
-  const skipMediaExtraction = options.skipMediaExtraction || false;
-
-  // Extract tables directly from Word XML (reliable, no heuristics)
-  const wordTables = await extractWordTables(docxPath);
-
-  // Try pandoc first with --track-changes=all to preserve reviewer edits
-  try {
-    // Build pandoc command
-    let pandocCmd = `pandoc "${docxPath}" -t markdown --wrap=none --track-changes=all`;
-    if (!skipMediaExtraction) {
-      pandocCmd += ` --extract-media="${mediaDir}"`;
-    }
-
-    const { stdout } = await execAsync(pandocCmd, { maxBuffer: 50 * 1024 * 1024 });
-    text = stdout;
-
-    // Convert pandoc's track change format to CriticMarkup
-    const origLength = text.length;
-
-    // Use a more robust pattern that handles nested content
-    text = text.replace(/\[([^\[\]]*(?:\[[^\[\]]*\][^\[\]]*)*)\]\{\.insertion[^}]*\}/g, (match, content) => {
-      if (content.trim()) {
-        trackChangeStats.insertions++;
-        return `{++${content}++}`;
-      }
-      return ''; // Empty insertions are removed
-    });
-
-    text = text.replace(/\[([^\[\]]*(?:\[[^\[\]]*\][^\[\]]*)*)\]\{\.deletion[^}]*\}/g, (match, content) => {
-      if (content.trim()) {
-        trackChangeStats.deletions++;
-        return `{--${content}--}`;
-      }
-      return ''; // Empty deletions are removed
-    });
-
-    // Handle any remaining pandoc track change patterns
-    let prevText;
-    do {
-      prevText = text;
-      text = text.replace(/\[([^\]]*)\]\{\.insertion[^}]*\}/g, (match, content) => {
-        if (content.trim()) {
-          trackChangeStats.insertions++;
-          return `{++${content}++}`;
-        }
-        return '';
-      });
-      text = text.replace(/\[([^\]]*)\]\{\.deletion[^}]*\}/g, (match, content) => {
-        if (content.trim()) {
-          trackChangeStats.deletions++;
-          return `{--${content}--}`;
-        }
-        return '';
-      });
-    } while (text !== prevText);
-
-    // Handle pandoc comment patterns - remove comment text from body
-    text = text.replace(/\[[^\]]*\]\{\.comment-start[^}]*\}/g, '');
-    text = text.replace(/\[\]\{\.comment-end[^}]*\}/g, '');
-
-    // Also handle {.mark} spans
-    text = text.replace(/\[([^\]]*)\]\{\.mark\}/g, '$1');
-
-    hasTrackChanges = trackChangeStats.insertions > 0 || trackChangeStats.deletions > 0;
-
-    if (hasTrackChanges) {
-      messages.push({
-        type: 'info',
-        message: `Found ${trackChangeStats.insertions} insertion(s) and ${trackChangeStats.deletions} deletion(s) from track changes`
-      });
-    }
-
-    // Find extracted media files
-    const mediaSubdir = path.join(mediaDir, 'media');
-    if (fs.existsSync(mediaSubdir)) {
-      extractedMedia = fs.readdirSync(mediaSubdir)
-        .filter(f => /\.(png|jpg|jpeg|gif|svg|emf|wmf|tiff?)$/i.test(f))
-        .map(f => path.join(mediaSubdir, f));
-
-      if (extractedMedia.length > 0) {
-        messages.push({
-          type: 'info',
-          message: `Extracted ${extractedMedia.length} image(s) to ${mediaSubdir}`
-        });
-      }
-    }
-  } catch (pandocErr: any) {
-    // Pandoc not available — use XML-based extraction with track change support
-    const { extractPlainTextWithTrackChanges } = await import('./word.js');
-    const { getInstallInstructions } = await import('./dependencies.js');
-    const installCmd = getInstallInstructions('pandoc');
-
-    const xmlResult = await extractPlainTextWithTrackChanges(docxPath);
-    text = xmlResult.text;
-    hasTrackChanges = xmlResult.hasTrackChanges;
-    trackChangeStats = xmlResult.stats;
-
-    if (hasTrackChanges) {
-      messages.push({
-        type: 'warning',
-        message: `Pandoc not installed. Using built-in XML extractor (${trackChangeStats.insertions} insertions, ${trackChangeStats.deletions} deletions preserved). Formatting may differ. Install pandoc for best results: ${installCmd}`
-      });
-    } else {
-      messages.push({
-        type: 'warning',
-        message: `Pandoc not installed. Using built-in XML extractor (no track changes found). Install pandoc for better formatting: ${installCmd}`
-      });
-    }
-  }
-
-  // Extract comments directly from docx XML
-  const comments = await extractWordComments(docxPath);
-
-  // Extract comment anchor texts
-  const { anchors } = await extractCommentAnchors(docxPath);
-
-  return {
-    text,
-    comments,
-    anchors,
-    messages,
-    extractedMedia,
-    tables: wordTables,
-    hasTrackChanges,
-    trackChangeStats,
-  };
-}
+/**
+ * Word document data extraction - raw extraction from .docx files
+ */
+
+import * as fs from 'fs';
+import * as path from 'path';
+import { exec } from 'child_process';
+import { promisify } from 'util';
+
+const execAsync = promisify(exec);
+
+// ============================================
+// Type Definitions
+// ============================================
+
+export interface WordComment {
+  id: string;
+  author: string;
+  date: string;
+  text: string;
+  /**
+   * Parent comment id when this is a reply in a Word comment thread.
+   * Resolved from `commentsExtended.xml`'s `w15:paraIdParent` field.
+   * `undefined` for top-level comments.
+   */
+  parentId?: string;
+}
+
+export interface TextNode {
+  xmlStart: number;
+  xmlEnd: number;
+  textStart: number;
+  textEnd: number;
+  text: string;
+}
+
+export interface CommentAnchorData {
+  anchor: string;
+  before: string;
+  after: string;
+  docPosition: number;
+  docLength: number;
+  isEmpty: boolean;
+}
+
+export interface CommentAnchorsResult {
+  anchors: Map<string, CommentAnchorData>;
+  fullDocText: string;
+}
+
+export interface DocxHeading {
+  /** Heading style name from `<w:pStyle>`, e.g. "Heading1" */
+  style: string;
+  /** Heading depth: 1, 2, 3, ... (parsed from style name; 0 if unknown) */
+  level: number;
+  /** Concatenated text content of the heading paragraph */
+  text: string;
+  /** Position in fullDocText (same coordinate system as CommentAnchorData.docPosition) */
+  docPosition: number;
+}
+
+export interface WordTable {
+  markdown: string;
+  rowCount: number;
+  colCount: number;
+}
+
+export interface ParsedRow {
+  cells: string[];
+  colSpans: number[];
+}
+
+export interface ExtractFromWordOptions {
+  mediaDir?: string;
+  skipMediaExtraction?: boolean;
+}
+
+export interface ExtractMessage {
+  type: 'info' | 'warning';
+  message: string;
+}
+
+export interface ExtractFromWordResult {
+  text: string;
+  comments: WordComment[];
+  anchors: Map<string, CommentAnchorData>;
+  messages: ExtractMessage[];
+  extractedMedia: string[];
+  tables: WordTable[];
+  hasTrackChanges: boolean;
+  trackChangeStats: { insertions: number; deletions: number };
+}
+
+// ============================================
+// Functions
+// ============================================
+
+/**
+ * Extract comments directly from Word docx comments.xml
+ */
+export async function extractWordComments(docxPath: string): Promise<WordComment[]> {
+  const AdmZip = (await import('adm-zip')).default;
+  const { parseStringPromise } = await import('xml2js');
+
+  const comments: WordComment[] = [];
+
+  // Validate file exists
+  if (!fs.existsSync(docxPath)) {
+    throw new Error(`File not found: ${docxPath}`);
+  }
+
+  try {
+    let zip;
+    try {
+      zip = new AdmZip(docxPath);
+    } catch (err: any) {
+      throw new Error(`Invalid Word document (not a valid .docx file): ${err.message}`);
+    }
+
+    const commentsEntry = zip.getEntry('word/comments.xml');
+
+    if (!commentsEntry) {
+      return comments;
+    }
+
+    let commentsXml;
+    try {
+      commentsXml = commentsEntry.getData().toString('utf8');
+    } catch (err: any) {
+      throw new Error(`Failed to read comments from document: ${err.message}`);
+    }
+
+    const parsed = await parseStringPromise(commentsXml, { explicitArray: false });
+
+    const commentsRoot = parsed['w:comments'];
+    if (!commentsRoot || !commentsRoot['w:comment']) {
+      return comments;
+    }
+
+    // Ensure it's an array
+    const commentNodes = Array.isArray(commentsRoot['w:comment'])
+      ? commentsRoot['w:comment']
+      : [commentsRoot['w:comment']];
+
+    // Map every paraId that lives inside a comment back to that comment's id.
+    // Word's commentsExtended.xml expresses threading via w15:paraIdParent,
+    // which references the parent's first <w:p>. Replies use a secondary
+    // (often-empty) <w:p>, so each comment may contribute multiple paraIds.
+    const paraIdToCommentId = new Map<string, string>();
+
+    for (const comment of commentNodes) {
+      const id = comment.$?.['w:id'] || '';
+      const author = comment.$?.['w:author'] || 'Unknown';
+      const date = comment.$?.['w:date'] || '';
+
+      // Extract text from nested w:p/w:r/w:t elements and record paraIds.
+      let text = '';
+      const extractText = (node: any): void => {
+        if (!node) return;
+        if (typeof node === 'string') {
+          text += node;
+          return;
+        }
+        if (node['w:t']) {
+          const t = node['w:t'];
+          text += typeof t === 'string' ? t : (t._ || t);
+        }
+        if (node['w:r']) {
+          const runs = Array.isArray(node['w:r']) ? node['w:r'] : [node['w:r']];
+          runs.forEach(extractText);
+        }
+        if (node['w:p']) {
+          const paras = Array.isArray(node['w:p']) ? node['w:p'] : [node['w:p']];
+          for (const para of paras) {
+            const paraId = para?.$?.['w14:paraId'];
+            if (paraId && id) paraIdToCommentId.set(paraId, id);
+            extractText(para);
+          }
+        }
+      };
+      extractText(comment);
+
+      comments.push({ id, author, date: date.slice(0, 10), text: text.trim() });
+    }
+
+    // Resolve parent links from commentsExtended.xml. Missing entry just
+    // means the docx has no threading metadata (e.g. legacy/non-Word source).
+    const extendedEntry = zip.getEntry('word/commentsExtended.xml');
+    if (extendedEntry && paraIdToCommentId.size > 0) {
+      let extendedXml = '';
+      try {
+        extendedXml = extendedEntry.getData().toString('utf8');
+      } catch {
+        // Unreadable threading metadata is non-fatal; skip parent linking.
+      }
+      if (extendedXml) {
+        const parentByCommentId = new Map<string, string>();
+        const exPattern = /<w15:commentEx\b([^>]*?)\/>/g;
+        let m: RegExpExecArray | null;
+        while ((m = exPattern.exec(extendedXml)) !== null) {
+          const attrs = m[1] ?? '';
+          const paraIdMatch = attrs.match(/w15:paraId="([^"]+)"/);
+          const parentMatch = attrs.match(/w15:paraIdParent="([^"]+)"/);
+          if (!paraIdMatch || !parentMatch) continue;
+          const childCommentId = paraIdToCommentId.get(paraIdMatch[1]);
+          const parentCommentId = paraIdToCommentId.get(parentMatch[1]);
+          if (childCommentId && parentCommentId && childCommentId !== parentCommentId) {
+            parentByCommentId.set(childCommentId, parentCommentId);
+          }
+        }
+        for (const c of comments) {
+          const parent = parentByCommentId.get(c.id);
+          if (parent) c.parentId = parent;
+        }
+      }
+    }
+  } catch (err: any) {
+    // Re-throw with more context if it's already an Error we created
+    if (err.message.includes('Invalid Word document') || err.message.includes('File not found')) {
+      throw err;
+    }
+    throw new Error(`Error extracting comments from ${path.basename(docxPath)}: ${err.message}`);
+  }
+
+  return comments;
+}
+
+/**
+ * Extract comment anchor texts from document.xml with surrounding context
+ * Returns map of comment ID -> {anchor, before, after, docPosition, isEmpty} for better matching
+ * Also returns fullDocText for section boundary matching
+ */
+export async function extractCommentAnchors(docxPath: string): Promise<CommentAnchorsResult> {
+  const AdmZip = (await import('adm-zip')).default;
+  const anchors = new Map<string, CommentAnchorData>();
+  let fullDocText = '';
+
+  try {
+    const zip = new AdmZip(docxPath);
+    const docEntry = zip.getEntry('word/document.xml');
+
+    if (!docEntry) {
+      return { anchors, fullDocText };
+    }
+
+    const docXml = docEntry.getData().toString('utf8');
+
+    // ========================================
+    // STEP 1: Build text position mapping
+    // ========================================
+    const textNodePattern = /<w:t[^>]*>([^<]*)<\/w:t>/g;
+    const textNodes: TextNode[] = [];
+    let textPosition = 0;
+    let nodeMatch;
+
+    while ((nodeMatch = textNodePattern.exec(docXml)) !== null) {
+      const rawText = nodeMatch[1] ?? '';
+      const decodedText = decodeXmlEntities(rawText);
+      textNodes.push({
+        xmlStart: nodeMatch.index,
+        xmlEnd: nodeMatch.index + nodeMatch[0].length,
+        textStart: textPosition,
+        textEnd: textPosition + decodedText.length,
+        text: decodedText
+      });
+      textPosition += decodedText.length;
+    }
+
+    fullDocText = textNodes.map(n => n.text).join('');
+
+    // Helper: convert XML position to text position
+    function xmlPosToTextPos(xmlPos: number): number {
+      for (let i = 0; i < textNodes.length; i++) {
+        const node = textNodes[i];
+        if (!node) continue;
+        if (xmlPos >= node.xmlStart && xmlPos < node.xmlEnd) {
+          return node.textStart;
+        }
+        if (xmlPos < node.xmlStart) {
+          return node.textStart;
+        }
+      }
+      const lastNode = textNodes[textNodes.length - 1];
+      return lastNode ? lastNode.textEnd : 0;
+    }
+
+    // Helper: extract context before a position
+    function getContextBefore(position: number, maxLength: number = 150): string {
+      const beforeText = fullDocText.slice(Math.max(0, position - maxLength), position);
+      const sentenceStart = beforeText.search(/[.!?]\s+[A-Z][^.!?]*$/);
+      return sentenceStart >= 0
+        ? beforeText.slice(sentenceStart + 2).trim()
+        : beforeText.slice(-80).trim();
+    }
+
+    // Helper: extract context after a position
+    function getContextAfter(position: number, maxLength: number = 150): string {
+      const afterText = fullDocText.slice(position, position + maxLength);
+      const sentenceEnd = afterText.search(/[.!?]\s/);
+      return sentenceEnd >= 0
+        ? afterText.slice(0, sentenceEnd + 1).trim()
+        : afterText.slice(0, 80).trim();
+    }
+
+    // ========================================
+    // STEP 2: Collect all start/end markers separately
+    // ========================================
+    const startPattern = /<w:commentRangeStart[^>]*w:id="(\d+)"[^>]*\/?>/g;
+    const endPattern = /<w:commentRangeEnd[^>]*w:id="(\d+)"[^>]*\/?>/g;
+
+    const starts = new Map<string, number>();  // id -> position after start tag
+    const ends = new Map<string, number>();    // id -> position before end tag
+
+    let match;
+    while ((match = startPattern.exec(docXml)) !== null) {
+      const id = match[1];
+      if (!starts.has(id)) {
+        starts.set(id, match.index + match[0].length);
+      }
+    }
+
+    while ((match = endPattern.exec(docXml)) !== null) {
+      const id = match[1];
+      if (!ends.has(id)) {
+        ends.set(id, match.index);
+      }
+    }
+
+    // ========================================
+    // STEP 3: Process each comment range by ID
+    // ========================================
+    for (const [id, startXmlPos] of starts) {
+      const endXmlPos = ends.get(id);
+
+      // Missing end marker - skip with warning
+      if (endXmlPos === undefined) {
+        console.warn(`Comment ${id}: missing end marker`);
+        continue;
+      }
+
+      // Calculate text position
+      const docPosition = xmlPosToTextPos(startXmlPos);
+
+      // Handle empty or inverted ranges
+      if (endXmlPos <= startXmlPos) {
+        anchors.set(id, {
+          anchor: '',
+          before: getContextBefore(docPosition),
+          after: getContextAfter(docPosition),
+          docPosition,
+          docLength: fullDocText.length,
+          isEmpty: true
+        });
+        continue;
+      }
+
+      // Extract XML segment between markers
+      const segment = docXml.slice(startXmlPos, endXmlPos);
+
+      // Extract text from w:t (regular) AND w:delText (deleted text in track changes)
+      const textInRangePattern = /<w:t[^>]*>([^<]*)<\/w:t>|<w:delText[^>]*>([^<]*)<\/w:delText>/g;
+      let anchorText = '';
+      let tm;
+      while ((tm = textInRangePattern.exec(segment)) !== null) {
+        anchorText += tm[1] || tm[2] || '';
+      }
+      anchorText = decodeXmlEntities(anchorText);
+
+      // Get context
+      const anchorLength = anchorText.length;
+      const before = getContextBefore(docPosition);
+      const after = getContextAfter(docPosition + anchorLength);
+
+      // ALWAYS add entry (even if anchor is empty)
+      anchors.set(id, {
+        anchor: anchorText.trim(),
+        before,
+        after,
+        docPosition,
+        docLength: fullDocText.length,
+        isEmpty: !anchorText.trim()
+      });
+    }
+  } catch (err: any) {
+    console.error('Error extracting comment anchors:', err.message);
+    return { anchors, fullDocText: '' };
+  }
+
+  return { anchors, fullDocText };
+}
+
+/**
+ * Extract heading paragraphs from a docx, with their text positions in the
+ * same coordinate system as `extractCommentAnchors`'s `fullDocText` and
+ * `CommentAnchorData.docPosition`.
+ *
+ * Headings are paragraphs whose `<w:pStyle>` is a Heading style. Reading
+ * styles directly is more reliable than keyword-matching the concatenated
+ * body text — there, paragraph boundaries are gone, so the literal string
+ * "Methods" can appear inside prose ("results across countries") and the
+ * structured-abstract label "Methods:" loses its colon when text runs are
+ * concatenated.
+ */
+export async function extractHeadings(docxPath: string): Promise<DocxHeading[]> {
+  const AdmZip = (await import('adm-zip')).default;
+
+  if (!fs.existsSync(docxPath)) {
+    throw new Error(`File not found: ${docxPath}`);
+  }
+
+  const zip = new AdmZip(docxPath);
+  const docEntry = zip.getEntry('word/document.xml');
+  if (!docEntry) return [];
+  const xml = docEntry.getData().toString('utf8');
+
+  // Build the same xml-pos → text-pos mapping that extractCommentAnchors does
+  const textNodePattern = /<w:t[^>]*>([^<]*)<\/w:t>/g;
+  const nodes: Array<{ xmlStart: number; xmlEnd: number; textStart: number; textEnd: number }> = [];
+  let textPos = 0;
+  let m;
+  while ((m = textNodePattern.exec(xml)) !== null) {
+    const decoded = decodeXmlEntities(m[1] ?? '');
+    nodes.push({
+      xmlStart: m.index,
+      xmlEnd: m.index + m[0].length,
+      textStart: textPos,
+      textEnd: textPos + decoded.length,
+    });
+    textPos += decoded.length;
+  }
+
+  function xmlToTextPos(xmlPos: number): number {
+    for (const n of nodes) {
+      if (xmlPos >= n.xmlStart && xmlPos < n.xmlEnd) return n.textStart;
+      if (xmlPos < n.xmlStart) return n.textStart;
+    }
+    return nodes.length ? nodes[nodes.length - 1].textEnd : 0;
+  }
+
+  const headings: DocxHeading[] = [];
+  const paraPattern = /<w:p\b[^>]*>([\s\S]*?)<\/w:p>/g;
+  let pm;
+  while ((pm = paraPattern.exec(xml)) !== null) {
+    const inner = pm[1];
+    const styleMatch = inner.match(/<w:pStyle[^>]*w:val="([^"]+)"/);
+    if (!styleMatch) continue;
+    const style = styleMatch[1];
+    if (!/heading/i.test(style)) continue;
+
+    // Concatenate text runs; include w:delText so a heading inside a tracked
+    // deletion is still surfaced (verifying anchors against an original draft)
+    const textInRange = /<w:t[^>]*>([^<]*)<\/w:t>|<w:delText[^>]*>([^<]*)<\/w:delText>/g;
+    let txt = '';
+    let tm;
+    while ((tm = textInRange.exec(inner)) !== null) {
+      txt += decodeXmlEntities(tm[1] || tm[2] || '');
+    }
+    const trimmed = txt.trim();
+    if (!trimmed) continue;
+
+    const levelMatch = style.match(/(\d+)/);
+    const level = levelMatch ? parseInt(levelMatch[1], 10) : 0;
+    headings.push({
+      style,
+      level,
+      text: trimmed,
+      docPosition: xmlToTextPos(pm.index),
+    });
+  }
+
+  return headings;
+}
+
+/**
+ * Decode XML entities in text
+ */
+function decodeXmlEntities(text: string): string {
+  return text
+    .replace(/&amp;/g, '&')
+    .replace(/&lt;/g, '<')
+    .replace(/&gt;/g, '>')
+    .replace(/&quot;/g, '"')
+    .replace(/&apos;/g, "'")
+    .replace(/&#(\d+);/g, (_, code) => String.fromCharCode(parseInt(code, 10)))
+    .replace(/&#x([0-9a-fA-F]+);/g, (_, code) => String.fromCharCode(parseInt(code, 16)));
+}
+
+/**
+ * Extract text content from a Word XML cell
+ */
+function extractCellText(cellXml: string): string {
+  const parts: string[] = [];
+
+  // Check for OMML math - replace with [math] placeholder
+  if (cellXml.includes('<m:oMath')) {
+    // Try to extract the text representation of math
+    const mathTextMatches = cellXml.match(/<m:t>([^<]*)<\/m:t>/g) || [];
+    if (mathTextMatches.length > 0) {
+      const mathText = mathTextMatches.map((t) => t.replace(/<[^>]+>/g, '')).join('');
+      parts.push(mathText);
+    } else {
+      parts.push('[math]');
+    }
+  }
+
+  // Extract regular text from w:t elements
+  const textMatches = cellXml.match(/<w:t[^>]*>([^<]*)<\/w:t>/g) || [];
+  for (const match of textMatches) {
+    const text = match.replace(/<[^>]+>/g, '');
+    if (text) {
+      parts.push(text);
+    }
+  }
+
+  let result = parts.join('').trim();
+  result = decodeXmlEntities(result);
+
+  // Escape pipe characters in cell content (would break table)
+  result = result.replace(/\|/g, '\\|');
+
+  return result;
+}
+
+/**
+ * Parse a table row, handling merged cells (gridSpan)
+ */
+function parseTableRow(rowXml: string, expectedCols: number): ParsedRow {
+  // Match cells - handle both <w:tc> and <w:tc ...>
+  const cellMatches = rowXml.match(/<w:tc(?:\s[^>]*)?>[\s\S]*?<\/w:tc>/g) || [];
+  const cells: string[] = [];
+  const colSpans: number[] = [];
+
+  for (const cellXml of cellMatches) {
+    // Check for horizontal merge (gridSpan)
+    const gridSpanMatch = cellXml.match(/<w:gridSpan\s+w:val="(\d+)"/);
+    const span = gridSpanMatch ? parseInt(gridSpanMatch[1], 10) : 1;
+
+    // Check for vertical merge continuation (vMerge without restart)
+    // If vMerge is present without w:val="restart", it's a continuation - use empty
+    const vMergeMatch = cellXml.match(/<w:vMerge(?:\s+w:val="([^"]+)")?/);
+    const isVMergeContinuation = vMergeMatch && vMergeMatch[1] !== 'restart';
+
+    const cellText = isVMergeContinuation ? '' : extractCellText(cellXml);
+
+    // Add the cell content
+    cells.push(cellText);
+    colSpans.push(span);
+
+    // For gridSpan > 1, add empty cells to maintain column alignment
+    for (let i = 1; i < span; i++) {
+      cells.push('');
+      colSpans.push(0); // 0 indicates this is a spanned cell
+    }
+  }
+
+  return { cells, colSpans };
+}
+
+/**
+ * Determine table grid column count from table XML
+ */
+function getTableGridCols(tableXml: string): number {
+  // Try to get from tblGrid
+  const gridColMatches = tableXml.match(/<w:gridCol/g) || [];
+  if (gridColMatches.length > 0) {
+    return gridColMatches.length;
+  }
+
+  // Fallback: count max cells in any row
+  const rowMatches = tableXml.match(/<w:tr[\s\S]*?<\/w:tr>/g) || [];
+  let maxCols = 0;
+  for (const rowXml of rowMatches) {
+    const { cells } = parseTableRow(rowXml, 0);
+    maxCols = Math.max(maxCols, cells.length);
+  }
+  return maxCols;
+}
+
+/**
+ * Extract tables directly from Word document XML and convert to markdown pipe tables
+ */
+export async function extractWordTables(docxPath: string): Promise<WordTable[]> {
+  const AdmZip = (await import('adm-zip')).default;
+  const tables: WordTable[] = [];
+
+  try {
+    const zip = new AdmZip(docxPath);
+    const docEntry = zip.getEntry('word/document.xml');
+
+    if (!docEntry) {
+      return tables;
+    }
+
+    const xml = docEntry.getData().toString('utf8');
+
+    // Find all table elements
+    const tableMatches = xml.match(/<w:tbl>[\s\S]*?<\/w:tbl>/g) || [];
+
+    for (const tableXml of tableMatches) {
+      // Determine expected column count from grid
+      const expectedCols = getTableGridCols(tableXml);
+
+      // Extract rows
+      const rowMatches = tableXml.match(/<w:tr[\s\S]*?<\/w:tr>/g) || [];
+      const rows: string[][] = [];
+
+      for (const rowXml of rowMatches) {
+        const { cells } = parseTableRow(rowXml, expectedCols);
+        if (cells.length > 0) {
+          rows.push(cells);
+        }
+      }
+
+      if (rows.length > 0) {
+        // Convert to markdown pipe table
+        const markdown = convertRowsToMarkdownTable(rows);
+        tables.push({ markdown, rowCount: rows.length, colCount: expectedCols || rows[0]?.length || 0 });
+      }
+    }
+  } catch (err: any) {
+    console.error('Error extracting tables from Word:', err.message);
+  }
+
+  return tables;
+}
+
+/**
+ * Convert array of rows (each row is array of cell strings) to markdown pipe table
+ */
+function convertRowsToMarkdownTable(rows: string[][]): string {
+  if (rows.length === 0) return '';
+
+  // Normalize column count (use max across all rows)
+  const colCount = Math.max(...rows.map((r) => r.length));
+
+  // Pad rows to have consistent column count
+  const normalizedRows = rows.map((row) => {
+    while (row.length < colCount) {
+      row.push('');
+    }
+    return row;
+  });
+
+  // Build markdown table
+  const lines: string[] = [];
+
+  // Header row
+  const header = normalizedRows[0];
+  lines.push('| ' + header.join(' | ') + ' |');
+
+  // Separator row
+  lines.push('|' + header.map(() => '---').join('|') + '|');
+
+  // Data rows
+  for (let i = 1; i < normalizedRows.length; i++) {
+    lines.push('| ' + normalizedRows[i].join(' | ') + ' |');
+  }
+
+  return lines.join('\n');
+}
+
+/**
+ * Extract text from Word document using pandoc with track changes preserved
+ */
+export async function extractFromWord(
+  docxPath: string,
+  options: ExtractFromWordOptions = {}
+): Promise<ExtractFromWordResult> {
+  let text: string;
+  let messages: ExtractMessage[] = [];
+  let extractedMedia: string[] = [];
+  let hasTrackChanges = false;
+  let trackChangeStats = { insertions: 0, deletions: 0 };
+
+  // Determine media extraction directory
+  const docxDir = path.dirname(docxPath);
+  const mediaDir = options.mediaDir || path.join(docxDir, 'media');
+
+  // Skip media extraction if figures already exist (e.g., when re-importing with existing source)
+  const skipMediaExtraction = options.skipMediaExtraction || false;
+
+  // Extract tables directly from Word XML (reliable, no heuristics)
+  const wordTables = await extractWordTables(docxPath);
+
+  // Try pandoc first with --track-changes=all to preserve reviewer edits
+  try {
+    // Build pandoc command
+    let pandocCmd = `pandoc "${docxPath}" -t markdown --wrap=none --track-changes=all`;
+    if (!skipMediaExtraction) {
+      pandocCmd += ` --extract-media="${mediaDir}"`;
+    }
+
+    const { stdout } = await execAsync(pandocCmd, { maxBuffer: 50 * 1024 * 1024 });
+    text = stdout;
+
+    // Convert pandoc's track change format to CriticMarkup
+    const origLength = text.length;
+
+    // Use a more robust pattern that handles nested content
+    text = text.replace(/\[([^\[\]]*(?:\[[^\[\]]*\][^\[\]]*)*)\]\{\.insertion[^}]*\}/g, (match, content) => {
+      if (content.trim()) {
+        trackChangeStats.insertions++;
+        return `{++${content}++}`;
+      }
+      return ''; // Empty insertions are removed
+    });
+
+    text = text.replace(/\[([^\[\]]*(?:\[[^\[\]]*\][^\[\]]*)*)\]\{\.deletion[^}]*\}/g, (match, content) => {
+      if (content.trim()) {
+        trackChangeStats.deletions++;
+        return `{--${content}--}`;
+      }
+      return ''; // Empty deletions are removed
+    });
+
+    // Handle any remaining pandoc track change patterns
+    let prevText;
+    do {
+      prevText = text;
+      text = text.replace(/\[([^\]]*)\]\{\.insertion[^}]*\}/g, (match, content) => {
+        if (content.trim()) {
+          trackChangeStats.insertions++;
+          return `{++${content}++}`;
+        }
+        return '';
+      });
+      text = text.replace(/\[([^\]]*)\]\{\.deletion[^}]*\}/g, (match, content) => {
+        if (content.trim()) {
+          trackChangeStats.deletions++;
+          return `{--${content}--}`;
+        }
+        return '';
+      });
+    } while (text !== prevText);
+
+    // Handle pandoc comment patterns - remove comment text from body
+    text = text.replace(/\[[^\]]*\]\{\.comment-start[^}]*\}/g, '');
+    text = text.replace(/\[\]\{\.comment-end[^}]*\}/g, '');
+
+    // Also handle {.mark} spans
+    text = text.replace(/\[([^\]]*)\]\{\.mark\}/g, '$1');
+
+    hasTrackChanges = trackChangeStats.insertions > 0 || trackChangeStats.deletions > 0;
+
+    if (hasTrackChanges) {
+      messages.push({
+        type: 'info',
+        message: `Found ${trackChangeStats.insertions} insertion(s) and ${trackChangeStats.deletions} deletion(s) from track changes`
+      });
+    }
+
+    // Find extracted media files
+    const mediaSubdir = path.join(mediaDir, 'media');
+    if (fs.existsSync(mediaSubdir)) {
+      extractedMedia = fs.readdirSync(mediaSubdir)
+        .filter(f => /\.(png|jpg|jpeg|gif|svg|emf|wmf|tiff?)$/i.test(f))
+        .map(f => path.join(mediaSubdir, f));
+
+      if (extractedMedia.length > 0) {
+        messages.push({
+          type: 'info',
+          message: `Extracted ${extractedMedia.length} image(s) to ${mediaSubdir}`
+        });
+      }
+    }
+  } catch (pandocErr: any) {
+    // Pandoc not available — use XML-based extraction with track change support
+    const { extractPlainTextWithTrackChanges } = await import('./word.js');
+    const { getInstallInstructions } = await import('./dependencies.js');
+    const installCmd = getInstallInstructions('pandoc');
+
+    const xmlResult = await extractPlainTextWithTrackChanges(docxPath);
+    text = xmlResult.text;
+    hasTrackChanges = xmlResult.hasTrackChanges;
+    trackChangeStats = xmlResult.stats;
+
+    if (hasTrackChanges) {
+      messages.push({
+        type: 'warning',
+        message: `Pandoc not installed. Using built-in XML extractor (${trackChangeStats.insertions} insertions, ${trackChangeStats.deletions} deletions preserved). Formatting may differ. Install pandoc for best results: ${installCmd}`
+      });
+    } else {
+      messages.push({
+        type: 'warning',
+        message: `Pandoc not installed. Using built-in XML extractor (no track changes found). Install pandoc for better formatting: ${installCmd}`
+      });
+    }
+  }
+
+  // Extract comments directly from docx XML
+  const comments = await extractWordComments(docxPath);
+
+  // Extract comment anchor texts
+  const { anchors } = await extractCommentAnchors(docxPath);
+
+  return {
+    text,
+    comments,
+    anchors,
+    messages,
+    extractedMedia,
+    tables: wordTables,
+    hasTrackChanges,
+    trackChangeStats,
+  };
+}