diff --git a/src/islands/documents/PdfToDocx.tsx b/src/islands/documents/PdfToDocx.tsx
index 3d564e9..72fbac9 100644
--- a/src/islands/documents/PdfToDocx.tsx
+++ b/src/islands/documents/PdfToDocx.tsx
@@ -4,7 +4,7 @@ import { Dropzone } from '@/components/ui/Dropzone';
 import { Alert } from '@/components/ui/Alert';
 import { ProgressBar } from '@/components/ui/ProgressBar';
 import { downloadService } from '@/services/download.service';
-import { reconstruct, textDensity, type TextItem, type DocParagraph } from '@/tools/documents/pdf-docx.lib';
+import { reconstruct, reconstructBlocks, textDensity, type TextItem, type DocBlock } from '@/tools/documents/pdf-docx.lib';
 import type { Lang } from '@/i18n/config';
 
 const OCR_MIN_CHARS = 8; // pages with fewer real characters are treated as scanned
@@ -21,7 +21,7 @@ const TR: Record<Lang, {
     forceOcr: 'Force OCR (scanned PDFs)', forceOcrHint: 'Run OCR on every page instead of only pages with no selectable text.',
     reading: (p, n) => `Reading text — page ${p} of ${n}…`, ocr: (p, n) => `Reading (OCR) — page ${p} of ${n}…`, building: 'Building the Word document…',
     another: 'Convert another',
-    note: 'The .docx contains editable, reflowable text (paragraphs and headings), not a pixel-perfect copy of the PDF layout. Exact positioning, tables and columns are not preserved.',
+    note: 'The .docx has editable text: headings, paragraphs and reconstructed tables. Table detection is best-effort — complex or borderless tables and multi-column layouts may still need cleanup.',
     errRead: 'Could not open this file — is it a valid PDF?', errConvert: 'Sorry, converting this PDF failed.',
   },
   id: {
@@ -31,7 +31,7 @@ const TR: Record<Lang, {
     forceOcr: 'Paksa OCR (PDF hasil pindaian)', forceOcrHint: 'Jalankan OCR pada setiap halaman, bukan hanya halaman tanpa teks yang dapat dipilih.',
     reading: (p, n) => `Membaca teks — halaman ${p} dari ${n}…`, ocr: (p, n) => `Membaca (OCR) — halaman ${p} dari ${n}…`, building: 'Menyusun dokumen Word…',
     another: 'Konversi yang lain',
-    note: 'Berkas .docx berisi teks yang dapat diedit dan disusun ulang (paragraf dan judul), bukan salinan tata letak PDF yang sempurna. Posisi persis, tabel, dan kolom tidak dipertahankan.',
+    note: 'Berkas .docx berisi teks yang dapat diedit: judul, paragraf, dan tabel yang direkonstruksi. Deteksi tabel bersifat best-effort — tabel rumit atau tanpa garis dan tata letak multi-kolom mungkin masih perlu dirapikan.',
     errRead: 'Tidak dapat membuka berkas ini — apakah PDF yang valid?', errConvert: 'Maaf, konversi PDF ini gagal.',
   },
 };
@@ -58,7 +58,7 @@ export default function PdfToDocx({ lang = 'en' }: { lang?: Lang }) {
     try {
       const pdf = await loadingTask.promise;
       const total = pdf.numPages;
-      const pages: DocParagraph[][] = [];
+      const pages: DocBlock[][] = [];
 
       for (let p = 1; p <= total; p++) {
         const page = await pdf.getPage(p);
@@ -77,7 +77,7 @@ export default function PdfToDocx({ lang = 'en' }: { lang?: Lang }) {
           pages.push(await ocrPage(page));
         } else {
           setStatus(t.reading(p, total));
-          pages.push(reconstruct(items));
+          pages.push(reconstructBlocks(items));
         }
         page.cleanup();
         setProgress(Math.round((p / total) * 90));
@@ -133,7 +133,7 @@ export default function PdfToDocx({ lang = 'en' }: { lang?: Lang }) {
 }
 
 // Render a page to a canvas and reconstruct paragraphs from on-device OCR.
-async function ocrPage(page: import('pdfjs-dist').PDFPageProxy): Promise<DocParagraph[]> {
+async function ocrPage(page: import('pdfjs-dist').PDFPageProxy): Promise<DocBlock[]> {
   const viewport = page.getViewport({ scale: 2 });
   const canvas = document.createElement('canvas');
   canvas.width = Math.floor(viewport.width);
@@ -147,21 +147,35 @@ async function ocrPage(page: import('pdfjs-dist').PDFPageProxy): Promise<DocPara
   const engine = await getEngine();
   const lines = await engine.recognize(canvas);
   const items: TextItem[] = lines.map((l) => ({ text: l.text, x: l.box.x, y: l.box.y, width: l.box.width, height: l.box.height }));
-  return reconstruct(items);
+  // OCR gives lines, not reliable column geometry, so keep the paragraph path.
+  return reconstruct(items).map((p) => ({ type: 'paragraph', ...p }));
 }
 
-async function buildDocx(pages: DocParagraph[][]): Promise<Blob> {
-  const { Document, Packer, Paragraph, HeadingLevel } = await import('docx');
-  const children: InstanceType<typeof Paragraph>[] = [];
-  pages.forEach((paras, pageIdx) => {
-    paras.forEach((par, i) => {
-      children.push(new Paragraph({
-        text: par.text,
-        heading: par.heading === 1 ? HeadingLevel.HEADING_1 : par.heading === 2 ? HeadingLevel.HEADING_2 : undefined,
-        pageBreakBefore: pageIdx > 0 && i === 0 ? true : undefined,
-      }));
+async function buildDocx(pages: DocBlock[][]): Promise<Blob> {
+  const { Document, Packer, Paragraph, HeadingLevel, Table, TableRow, TableCell, WidthType } = await import('docx');
+  const children: (InstanceType<typeof Paragraph> | InstanceType<typeof Table>)[] = [];
+
+  pages.forEach((blocks, pageIdx) => {
+    blocks.forEach((block, i) => {
+      const first = pageIdx > 0 && i === 0;
+      if (block.type === 'table') {
+        children.push(new Table({
+          width: { size: 100, type: WidthType.PERCENTAGE },
+          rows: block.rows.map((row) => new TableRow({
+            children: row.map((cell) => new TableCell({ children: [new Paragraph({ text: cell })] })),
+          })),
+        }));
+        children.push(new Paragraph({ text: '' })); // spacing after the table
+      } else {
+        children.push(new Paragraph({
+          text: block.text,
+          heading: block.heading === 1 ? HeadingLevel.HEADING_1 : block.heading === 2 ? HeadingLevel.HEADING_2 : undefined,
+          pageBreakBefore: first ? true : undefined,
+        }));
+      }
     });
   });
+
   if (children.length === 0) children.push(new Paragraph({ text: '' }));
   const doc = new Document({ sections: [{ children }] });
   return Packer.toBlob(doc);
diff --git a/src/tools/documents/pdf-docx.e2e.test.ts b/src/tools/documents/pdf-docx.e2e.test.ts
new file mode 100644
index 0000000..1fd5370
--- /dev/null
+++ b/src/tools/documents/pdf-docx.e2e.test.ts
@@ -0,0 +1,38 @@
+import { describe, it, expect } from 'vitest';
+import { reconstructBlocks, type TextItem } from './pdf-docx.lib';
+
+// Full pipeline: draw a PDF (pdf-lib) → extract positioned text (pdf.js) →
+// reconstruct blocks. Guards against regressions like pdf.js's synthetic
+// whitespace items merging table cells.
+describe('pdf → docx reconstruction (end to end)', () => {
+  it('reconstructs a drawn table as a real table, prose as paragraphs', async () => {
+    const { PDFDocument, StandardFonts } = await import('pdf-lib');
+    const doc = await PDFDocument.create();
+    const page = doc.addPage([612, 792]);
+    const font = await doc.embedFont(StandardFonts.Helvetica);
+    const draw = (text: string, x: number, y: number, size = 12) => page.drawText(text, { x, y, size, font });
+    draw('A short introductory paragraph of text on its own.', 50, 715);
+    const rows = [['Item', 'Q1', 'Q2'], ['Revenue', '100', '150'], ['Costs', '40', '55']];
+    const cx = [50, 260, 400];
+    rows.forEach((r, ri) => r.forEach((c, ci) => draw(c, cx[ci], 670 - ri * 22)));
+    const bytes = await doc.save();
+
+    const pdfjs = await import('pdfjs-dist/legacy/build/pdf.mjs');
+    const pdf = await pdfjs.getDocument({ data: bytes, isEvalSupported: false, useWorkerFetch: false }).promise;
+    const p = await pdf.getPage(1);
+    const viewport = p.getViewport({ scale: 1 });
+    const tc = await p.getTextContent();
+    const items: TextItem[] = tc.items
+      .filter((i): i is Extract<typeof i, { str: string }> => 'str' in i)
+      .map((i) => {
+        const tr = pdfjs.Util.transform(viewport.transform, i.transform);
+        return { text: i.str, x: tr[4], y: tr[5], width: i.width, height: Math.hypot(tr[2], tr[3]) || 10 };
+      });
+
+    const blocks = reconstructBlocks(items);
+    const table = blocks.find((b) => b.type === 'table');
+    expect(table).toBeDefined();
+    if (table && table.type === 'table') expect(table.rows).toEqual(rows);
+    expect(blocks.some((b) => b.type === 'paragraph' && b.text.includes('introductory'))).toBe(true);
+  });
+});
diff --git a/src/tools/documents/pdf-docx.lib.test.ts b/src/tools/documents/pdf-docx.lib.test.ts
index f331629..ead4e96 100644
--- a/src/tools/documents/pdf-docx.lib.test.ts
+++ b/src/tools/documents/pdf-docx.lib.test.ts
@@ -1,5 +1,5 @@
 import { describe, it, expect } from 'vitest';
-import { groupLines, paragraphsFromLines, reconstruct, textDensity, type TextItem } from './pdf-docx.lib';
+import { groupLines, paragraphsFromLines, reconstruct, textDensity, segmentsOf, clusterColumns, reconstructBlocks, type TextItem } from './pdf-docx.lib';
 
 const item = (text: string, x: number, y: number, width: number, height = 10): TextItem => ({ text, x, y, width, height });
 
@@ -69,3 +69,52 @@ describe('reconstruct + textDensity', () => {
     expect(textDensity([])).toBe(0);
   });
 });
+
+describe('segmentsOf', () => {
+  it('splits a row into cells on wide column gaps but keeps words together', () => {
+    // "Name Field" as one cell (small gap), then a big gap, then "Value"
+    const items = [
+      item('Name', 10, 100, 30), item('Field', 46, 100, 30),
+      item('Value', 200, 100, 40),
+    ];
+    const segs = segmentsOf(items);
+    expect(segs.map(s => s.text)).toEqual(['Name Field', 'Value']);
+  });
+});
+
+describe('clusterColumns', () => {
+  it('groups nearby x positions into column centres', () => {
+    expect(clusterColumns([10, 11, 100, 101, 200], 5)).toEqual([10.5, 100.5, 200]);
+  });
+});
+
+describe('reconstructBlocks', () => {
+  const cell = (text: string, x: number, y: number) => item(text, x, y, 40, 10);
+
+  it('emits a table for aligned multi-column rows', () => {
+    const items: TextItem[] = [
+      cell('Item', 10, 100), cell('Q1', 200, 100), cell('Q2', 300, 100),
+      cell('Rev', 10, 120), cell('100', 200, 120), cell('150', 300, 120),
+      cell('Cost', 10, 140), cell('40', 200, 140), cell('55', 300, 140),
+    ];
+    const blocks = reconstructBlocks(items);
+    const table = blocks.find(b => b.type === 'table');
+    expect(table).toBeDefined();
+    if (table && table.type === 'table') {
+      expect(table.rows).toEqual([
+        ['Item', 'Q1', 'Q2'],
+        ['Rev', '100', '150'],
+        ['Cost', '40', '55'],
+      ]);
+    }
+  });
+
+  it('keeps ordinary prose as paragraphs', () => {
+    const items: TextItem[] = [
+      item('This is a normal sentence of text.', 10, 100, 300, 10),
+      item('Another line right below it.', 10, 118, 260, 10),
+    ];
+    const blocks = reconstructBlocks(items);
+    expect(blocks.every(b => b.type === 'paragraph')).toBe(true);
+  });
+});
diff --git a/src/tools/documents/pdf-docx.lib.ts b/src/tools/documents/pdf-docx.lib.ts
index b8e0e13..ae23729 100644
--- a/src/tools/documents/pdf-docx.lib.ts
+++ b/src/tools/documents/pdf-docx.lib.ts
@@ -119,6 +119,144 @@ export function reconstruct(items: TextItem[]): DocParagraph[] {
   return paragraphsFromLines(groupLines(items));
 }
 
+/* ------------------------------------------------------------------ *
+ * Table-aware reconstruction: detect grid-like regions (rows whose text
+ * splits into columns that line up across several rows) and emit them as
+ * tables, with everything else falling back to paragraphs.
+ * ------------------------------------------------------------------ */
+
+export interface Segment { x: number; text: string; }
+interface ItemRow { y: number; height: number; segs: Segment[]; }
+
+export type DocBlock =
+  | { type: 'paragraph'; text: string; heading: 0 | 1 | 2 }
+  | { type: 'table'; rows: string[][] };
+
+/** Group items into rows, keeping each row's items (not just joined text). */
+function groupItemRows(items: TextItem[], yTolRatio = 0.5): TextItem[][] {
+  // Drop whitespace-only items — pdf.js emits synthetic space items between
+  // cells that would otherwise bridge column gaps and merge table cells.
+  const valid = items.filter((it) => it.text.trim().length > 0);
+  if (!valid.length) return [];
+  const medH = median(valid.map((i) => i.height).filter((h) => h > 0)) || 1;
+  const tol = medH * yTolRatio;
+  const sorted = [...valid].sort((a, b) => a.y - b.y || a.x - b.x);
+  const groups: TextItem[][] = [];
+  for (const it of sorted) {
+    const last = groups[groups.length - 1];
+    const lastY = last ? last.reduce((s, g) => s + g.y, 0) / last.length : 0;
+    if (last && Math.abs(it.y - lastY) <= tol) last.push(it);
+    else groups.push([it]);
+  }
+  return groups.map((g) => [...g].sort((a, b) => a.x - b.x));
+}
+
+/** Split a row's items into cell segments, breaking on wide (column) gaps. */
+export function segmentsOf(lineItems: TextItem[], gapFactor = 1): Segment[] {
+  const sorted = [...lineItems].sort((a, b) => a.x - b.x);
+  const segs: Segment[] = [];
+  let text = '';
+  let x = 0;
+  let prevEnd: number | null = null;
+  for (const it of sorted) {
+    if (it.text.trim().length === 0) continue; // skip synthetic whitespace items
+    if (prevEnd !== null && it.x - prevEnd > it.height * gapFactor) {
+      if (text.trim()) segs.push({ x, text: text.trim() });
+      text = '';
+    }
+    if (text === '') x = it.x;
+    else if (prevEnd !== null && it.x - prevEnd > it.height * 0.2 && !/\s$/.test(text)) text += ' ';
+    text += it.text;
+    prevEnd = it.x + it.width;
+  }
+  if (text.trim()) segs.push({ x, text: text.trim() });
+  return segs;
+}
+
+/** Cluster x positions into column centres (1-D agglomerative by tolerance). */
+export function clusterColumns(xs: number[], tol: number): number[] {
+  const sorted = [...xs].sort((a, b) => a - b);
+  const cols: number[][] = [];
+  for (const v of sorted) {
+    const last = cols[cols.length - 1];
+    const c = last ? last.reduce((s, n) => s + n, 0) / last.length : 0;
+    if (last && Math.abs(v - c) <= tol) last.push(v);
+    else cols.push([v]);
+  }
+  return cols.map((c) => c.reduce((s, n) => s + n, 0) / c.length);
+}
+
+const nearestCol = (x: number, cols: number[]): number => {
+  let best = 0;
+  let bestD = Infinity;
+  for (let i = 0; i < cols.length; i++) {
+    const d = Math.abs(x - cols[i]);
+    if (d < bestD) { bestD = d; best = i; }
+  }
+  return best;
+};
+
+/**
+ * Reconstruct a page's text items into a mix of paragraph and table blocks.
+ * A table is a run of ≥2 adjacent rows that each split into ≥2 aligned columns.
+ */
+export function reconstructBlocks(items: TextItem[]): DocBlock[] {
+  const itemRows = groupItemRows(items);
+  if (!itemRows.length) return [];
+  const medH = median(itemRows.flat().map((i) => i.height).filter((h) => h > 0)) || 1;
+
+  const rows: ItemRow[] = itemRows.map((g) => ({
+    y: g.reduce((s, i) => s + i.y, 0) / g.length,
+    height: Math.max(...g.map((i) => i.height)),
+    segs: segmentsOf(g),
+  }));
+
+  const blocks: DocBlock[] = [];
+  let paraLines: DocLine[] = [];
+  const flushParas = () => {
+    if (!paraLines.length) return;
+    for (const p of paragraphsFromLines(paraLines)) blocks.push({ type: 'paragraph', ...p });
+    paraLines = [];
+  };
+  const asLine = (r: ItemRow): DocLine => ({
+    text: r.segs.map((s) => s.text).join(' ').replace(/\s+/g, ' ').trim(),
+    x: r.segs.length ? r.segs[0].x : 0,
+    y: r.y,
+    height: r.height,
+  });
+
+  let i = 0;
+  while (i < rows.length) {
+    // A table candidate starts on a multi-column row.
+    if (rows[i].segs.length >= 2) {
+      let j = i + 1;
+      while (j < rows.length && rows[j].segs.length >= 2 && rows[j].y - rows[j - 1].y <= medH * 2.5) j++;
+      const region = rows.slice(i, j);
+      if (region.length >= 2) {
+        const cols = clusterColumns(region.flatMap((r) => r.segs.map((s) => s.x)), medH * 1.5);
+        if (cols.length >= 2) {
+          flushParas();
+          const grid = region.map((r) => {
+            const cells = Array<string>(cols.length).fill('');
+            for (const seg of r.segs) {
+              const c = nearestCol(seg.x, cols);
+              cells[c] = cells[c] ? `${cells[c]} ${seg.text}` : seg.text;
+            }
+            return cells;
+          });
+          blocks.push({ type: 'table', rows: grid });
+          i = j;
+          continue;
+        }
+      }
+    }
+    paraLines.push(asLine(rows[i]));
+    i++;
+  }
+  flushParas();
+  return blocks;
+}
+
 /** Count of non-whitespace characters — used to decide if a page needs OCR. */
 export function textDensity(items: TextItem[]): number {
   return items.reduce((n, it) => n + it.text.replace(/\s/g, '').length, 0);