@clazic/kordoc 2.3.1 → 2.4.0

package/dist/index.d.cts

@@ -1,3 +1,5 @@
+import JSZip from 'jszip';
+
 /** kordoc 공통 타입 정의 */
 interface CellContext {
     text: string;
@@ -141,8 +143,8 @@ interface OutlineItem {
     pageNumber?: number;
 }
 /** 구조화된 에러 코드 — 프로그래밍적 에러 핸들링용 */
-type ErrorCode = "EMPTY_INPUT" | "UNSUPPORTED_FORMAT" | "ENCRYPTED" | "DRM_PROTECTED" | "CORRUPTED" | "DECOMPRESSION_BOMB" | "ZIP_BOMB" | "IMAGE_BASED_PDF" | "NO_SECTIONS" | "PARSE_ERROR";
-type FileType = "hwpx" | "hwp" | "pdf" | "xlsx" | "docx" | "unknown";
+type ErrorCode = "EMPTY_INPUT" | "UNSUPPORTED_FORMAT" | "ENCRYPTED" | "DRM_PROTECTED" | "CORRUPTED" | "DECOMPRESSION_BOMB" | "ZIP_BOMB" | "IMAGE_BASED_PDF" | "NO_SECTIONS" | "PARSE_ERROR" | "FILE_TOO_LARGE";
+type FileType = "hwpx" | "hwp" | "pdf" | "xlsx" | "docx" | "image" | "unknown";
 interface ParseResultBase {
     fileType: FileType;
     /** 페이지/섹션 수 — PDF: 실제 페이지 수, HWP/HWPX: 섹션 수, XLSX: 시트 수 */
@@ -309,7 +311,10 @@ declare function detectFormat(buffer: ArrayBuffer): FileType;
  * ZIP 내부 구조 기반 포맷 세분화.
  * HWPX, XLSX, DOCX 모두 ZIP이므로 내부 파일로 구분.
  */
-declare function detectZipFormat(buffer: ArrayBuffer): Promise<"hwpx" | "xlsx" | "docx" | "unknown">;
+declare function detectZipFormat(buffer: ArrayBuffer): Promise<{
+    format: "hwpx" | "xlsx" | "docx" | "unknown";
+    zip: JSZip | null;
+}>;
 
 /** 2-pass colSpan/rowSpan 테이블 빌더 및 Markdown 변환 */
 
@@ -338,14 +343,14 @@ declare const VERSION: string;
  */
 declare function parse(input: string | ArrayBuffer | Buffer, options?: ParseOptions): Promise<ParseResult>;
 /** HWPX 파일을 Markdown으로 변환 */
-declare function parseHwpx(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
+declare function parseHwpx(buffer: ArrayBuffer, options?: ParseOptions, zip?: JSZip): Promise<ParseResult>;
 /** HWP 5.x 바이너리 파일을 Markdown으로 변환 */
 declare function parseHwp(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
 /** PDF 파일에서 텍스트를 추출하여 Markdown으로 변환 */
 declare function parsePdf(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
 /** XLSX 파일을 Markdown으로 변환 */
-declare function parseXlsx(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
+declare function parseXlsx(buffer: ArrayBuffer, options?: ParseOptions, zip?: JSZip): Promise<ParseResult>;
 /** DOCX 파일을 Markdown으로 변환 */
-declare function parseDocx(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
+declare function parseDocx(buffer: ArrayBuffer, options?: ParseOptions, zip?: JSZip): Promise<ParseResult>;
 
 export { type BlockDiff, type BoundingBox, type CellContext, type CellDiff, type DiffChangeType, type DiffResult, type DocumentMetadata, type ErrorCode, type ExtractedImage, type FileType, type FormField, type FormResult, type IRBlock, type IRBlockType, type IRCell, type IRTable, type ImageData, type InlineStyle, type MarkdownToXlsxOptions, type OcrMode, type OcrProvider, type OutlineItem, type ParseFailure, type ParseOptions, type ParseResult, type ParseSuccess, type ParseWarning, type StructuredOcrResult, VERSION, type WarningCode, type WatchOptions, blocksToMarkdown, compare, detectFormat, detectZipFormat, diffBlocks, extractFormFields, isHwpxFile, isOldHwpFile, isPdfFile, isZipFile, markdownToHwpx, markdownToXlsx, parse, parseDocx, parseHwp, parseHwpx, parsePdf, parseXlsx };

package/dist/index.d.ts

@@ -1,3 +1,5 @@
+import JSZip from 'jszip';
+
 /** kordoc 공통 타입 정의 */
 interface CellContext {
     text: string;
@@ -141,8 +143,8 @@ interface OutlineItem {
     pageNumber?: number;
 }
 /** 구조화된 에러 코드 — 프로그래밍적 에러 핸들링용 */
-type ErrorCode = "EMPTY_INPUT" | "UNSUPPORTED_FORMAT" | "ENCRYPTED" | "DRM_PROTECTED" | "CORRUPTED" | "DECOMPRESSION_BOMB" | "ZIP_BOMB" | "IMAGE_BASED_PDF" | "NO_SECTIONS" | "PARSE_ERROR";
-type FileType = "hwpx" | "hwp" | "pdf" | "xlsx" | "docx" | "unknown";
+type ErrorCode = "EMPTY_INPUT" | "UNSUPPORTED_FORMAT" | "ENCRYPTED" | "DRM_PROTECTED" | "CORRUPTED" | "DECOMPRESSION_BOMB" | "ZIP_BOMB" | "IMAGE_BASED_PDF" | "NO_SECTIONS" | "PARSE_ERROR" | "FILE_TOO_LARGE";
+type FileType = "hwpx" | "hwp" | "pdf" | "xlsx" | "docx" | "image" | "unknown";
 interface ParseResultBase {
     fileType: FileType;
     /** 페이지/섹션 수 — PDF: 실제 페이지 수, HWP/HWPX: 섹션 수, XLSX: 시트 수 */
@@ -309,7 +311,10 @@ declare function detectFormat(buffer: ArrayBuffer): FileType;
  * ZIP 내부 구조 기반 포맷 세분화.
  * HWPX, XLSX, DOCX 모두 ZIP이므로 내부 파일로 구분.
  */
-declare function detectZipFormat(buffer: ArrayBuffer): Promise<"hwpx" | "xlsx" | "docx" | "unknown">;
+declare function detectZipFormat(buffer: ArrayBuffer): Promise<{
+    format: "hwpx" | "xlsx" | "docx" | "unknown";
+    zip: JSZip | null;
+}>;
 
 /** 2-pass colSpan/rowSpan 테이블 빌더 및 Markdown 변환 */
 
@@ -338,14 +343,14 @@ declare const VERSION: string;
  */
 declare function parse(input: string | ArrayBuffer | Buffer, options?: ParseOptions): Promise<ParseResult>;
 /** HWPX 파일을 Markdown으로 변환 */
-declare function parseHwpx(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
+declare function parseHwpx(buffer: ArrayBuffer, options?: ParseOptions, zip?: JSZip): Promise<ParseResult>;
 /** HWP 5.x 바이너리 파일을 Markdown으로 변환 */
 declare function parseHwp(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
 /** PDF 파일에서 텍스트를 추출하여 Markdown으로 변환 */
 declare function parsePdf(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
 /** XLSX 파일을 Markdown으로 변환 */
-declare function parseXlsx(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
+declare function parseXlsx(buffer: ArrayBuffer, options?: ParseOptions, zip?: JSZip): Promise<ParseResult>;
 /** DOCX 파일을 Markdown으로 변환 */
-declare function parseDocx(buffer: ArrayBuffer, options?: ParseOptions): Promise<ParseResult>;
+declare function parseDocx(buffer: ArrayBuffer, options?: ParseOptions, zip?: JSZip): Promise<ParseResult>;
 
 export { type BlockDiff, type BoundingBox, type CellContext, type CellDiff, type DiffChangeType, type DiffResult, type DocumentMetadata, type ErrorCode, type ExtractedImage, type FileType, type FormField, type FormResult, type IRBlock, type IRBlockType, type IRCell, type IRTable, type ImageData, type InlineStyle, type MarkdownToXlsxOptions, type OcrMode, type OcrProvider, type OutlineItem, type ParseFailure, type ParseOptions, type ParseResult, type ParseSuccess, type ParseWarning, type StructuredOcrResult, VERSION, type WarningCode, type WatchOptions, blocksToMarkdown, compare, detectFormat, detectZipFormat, diffBlocks, extractFormFields, isHwpxFile, isOldHwpFile, isPdfFile, isZipFile, markdownToHwpx, markdownToXlsx, parse, parseDocx, parseHwp, parseHwpx, parsePdf, parseXlsx };

package/dist/index.js

@@ -1998,8 +1998,8 @@ function getTesseractFallbackMessage() {
     "\uC124\uCE58\uB41C AI CLI\uAC00 \uC5C6\uC5B4 \uB0B4\uC7A5 tesseract.js\uB85C OCR\uC744 \uC218\uD589\uD569\uB2C8\uB2E4.",
     "\uB354 \uB098\uC740 \uD488\uC9C8(\uD14C\uC774\uBE14/\uD5E4\uB529 \uAD6C\uC870 \uBCF4\uC874)\uC744 \uC704\uD574 AI CLI \uC124\uCE58\uB97C \uAD8C\uC7A5\uD569\uB2C8\uB2E4:",
     "",
-    "  [\uAD8C\uC7A5] Gemini CLI:  https://ai.google.dev/gemini-api/docs/cli",
-    "         Codex CLI:   npm install -g @openai/codex",
+    "  [\uAD8C\uC7A5] Codex CLI:   npm install -g @openai/codex",
+    "         Gemini CLI:  https://ai.google.dev/gemini-api/docs/cli",
     "         Claude CLI:  npm install -g @anthropic-ai/claude-code",
     "         Ollama:      brew install ollama (+ ollama pull gemma4:27b)"
   ].join("\n");
@@ -2008,7 +2008,7 @@ var CLI_PRIORITY;
 var init_auto_detect = __esm({
   "src/ocr/auto-detect.ts"() {
     "use strict";
-    CLI_PRIORITY = ["gemini", "codex", "claude", "ollama"];
+    CLI_PRIORITY = ["codex", "gemini", "claude", "ollama"];
   }
 });
 
@@ -2051,7 +2051,7 @@ function callCli(mode, imagePath) {
   const args = buildCliArgs(mode, imagePath);
   const result = spawnSync(mode, args, {
     encoding: "utf-8",
-    timeout: 18e4,
+    timeout: 6e5,
     maxBuffer: 10 * 1024 * 1024,
     // claude: /tmp에서 실행하여 프로젝트 CLAUDE.md의 규칙 간섭 방지
     ...mode === "claude" ? { cwd: tmpdir() } : {}
@@ -2145,14 +2145,22 @@ async function callOllamaApi(imagePath) {
   return data.message?.content || "";
 }
 function stripCodeFence(text) {
-  const match = text.match(/^```(?:markdown|md)?\s*\n([\s\S]*?)\n```\s*$/m);
+  const match = text.match(/^```(?:markdown|md)?\s*([\s\S]*?)```\s*$/m);
   return match ? match[1].trim() : text;
 }
 var OCR_PROMPT, _tempDir;
 var init_cli_provider = __esm({
   "src/ocr/cli-provider.ts"() {
     "use strict";
-    OCR_PROMPT = "\uC774 PDF \uD398\uC774\uC9C0 \uC774\uBBF8\uC9C0\uC5D0\uC11C \uD14D\uC2A4\uD2B8\uC640 \uD14C\uC774\uBE14\uC744 \uCD94\uCD9C\uD558\uC5EC \uC21C\uC218 Markdown\uC73C\uB85C \uBCC0\uD658\uD558\uC138\uC694.\n\uADDC\uCE59:\n- \uD14C\uC774\uBE14\uC740 Markdown \uD14C\uC774\uBE14 \uBB38\uBC95 \uC0AC\uC6A9 (| \uAD6C\uBD84, |---|---| \uD5E4\uB354 \uAD6C\uBD84\uC120 \uD3EC\uD568)\n- \uBCD1\uD569\uB41C \uC140\uC740 \uD574\uB2F9 \uC704\uCE58\uC5D0 \uB0B4\uC6A9 \uAE30\uC7AC\n- \uD5E4\uB529\uC740 \uAE00\uC790 \uD06C\uAE30\uC5D0 \uB530\uB77C ## ~ ###### \uC0AC\uC6A9\n- \uB9AC\uC2A4\uD2B8\uB294 - \uB610\uB294 1. \uC0AC\uC6A9\n- \uC774\uBBF8\uC9C0, \uB3C4\uD615 \uB4F1 \uBE44\uD14D\uC2A4\uD2B8 \uC694\uC18C\uB294 \uBB34\uC2DC\n- \uC6D0\uBB38\uC758 \uC77D\uAE30 \uC21C\uC11C\uC640 \uAD6C\uC870\uB97C \uC720\uC9C0\n- ```\uB85C \uAC10\uC2F8\uC9C0 \uB9D0\uACE0 \uC21C\uC218 Markdown\uB9CC \uCD9C\uB825";
+    OCR_PROMPT = `\uC774 PDF \uD398\uC774\uC9C0 \uC774\uBBF8\uC9C0\uC5D0\uC11C \uD14D\uC2A4\uD2B8\uC640 \uD14C\uC774\uBE14\uC744 \uCD94\uCD9C\uD558\uC5EC \uC21C\uC218 Markdown\uC73C\uB85C \uBCC0\uD658\uD558\uC138\uC694.
+\uADDC\uCE59:
+- \uD14C\uC774\uBE14\uC740 Markdown \uD14C\uC774\uBE14 \uBB38\uBC95 \uC0AC\uC6A9 (| \uAD6C\uBD84, |---|---| \uD5E4\uB354 \uAD6C\uBD84\uC120 \uD3EC\uD568)
+- \uBCD1\uD569\uB41C \uC140\uC740 \uD574\uB2F9 \uC704\uCE58\uC5D0 \uB0B4\uC6A9 \uAE30\uC7AC
+- \uD5E4\uB529\uC740 \uAE00\uC790 \uD06C\uAE30\uC5D0 \uB530\uB77C ## ~ ###### \uC0AC\uC6A9
+- \uB9AC\uC2A4\uD2B8\uB294 - \uB610\uB294 1. \uC0AC\uC6A9
+- \uC774\uBBF8\uC9C0, \uB3C4\uD615 \uB4F1 \uBE44\uD14D\uC2A4\uD2B8 \uC694\uC18C\uB294 \uBB34\uC2DC
+- \uC6D0\uBB38\uC758 \uC77D\uAE30 \uC21C\uC11C\uC640 \uAD6C\uC870\uB97C \uC720\uC9C0
+- \`\`\`\uB85C \uAC10\uC2F8\uC9C0 \uB9D0\uACE0 \uC21C\uC218 Markdown\uB9CC \uCD9C\uB825`;
     _tempDir = null;
   }
 });
@@ -2321,9 +2329,8 @@ async function callBatchCli(mode, imagePaths) {
 ${fileRefs}`;
   let args;
   if (mode === "gemini") {
-    args = ["--prompt", prompt, "--yolo"];
-    const model = process.env.KORDOC_GEMINI_MODEL;
-    if (model) args.push("--model", model);
+    const model = process.env.KORDOC_GEMINI_MODEL ?? "gemini-2.5-flash";
+    args = ["--prompt", prompt, "--yolo", "--model", model];
   } else {
     args = ["--print", prompt];
     const model = process.env.KORDOC_CLAUDE_MODEL;
@@ -2667,22 +2674,22 @@ async function ocrPagesBatch(doc, provider, pageFilter, effectivePageCount, warn
     if (pageFilter && !pageFilter.has(i)) continue;
     pageNumbers.push(i);
   }
-  const pageImages = [];
-  for (const pageNum of pageNumbers) {
-    const page = await doc.getPage(pageNum);
-    const image = await renderPageToPng(page);
-    pageImages.push({ image, pageNum });
-  }
-  const batches = [];
-  for (let i = 0; i < pageImages.length; i += provider.batchSize) {
-    batches.push(pageImages.slice(i, i + provider.batchSize));
+  const pageBatches = [];
+  for (let i = 0; i < pageNumbers.length; i += provider.batchSize) {
+    pageBatches.push(pageNumbers.slice(i, i + provider.batchSize));
   }
   let processed = 0;
-  const batchTasks = batches.map((batch, batchIdx) => async () => {
+  const batchTasks = pageBatches.map((batchPageNums, batchIdx) => async () => {
     const pageBlocks = [];
     try {
-      const results = await provider.processBatch(batch);
-      for (const { pageNum } of batch) {
+      const batchImages = [];
+      for (const pageNum of batchPageNums) {
+        const page = await doc.getPage(pageNum);
+        const image = await renderPageToPng(page);
+        batchImages.push({ image, pageNum });
+      }
+      const results = await provider.processBatch(batchImages);
+      for (const { pageNum } of batchImages) {
         const result = results.get(pageNum);
         pageBlocks.push({
           pageNum,
@@ -2690,16 +2697,16 @@ async function ocrPagesBatch(doc, provider, pageFilter, effectivePageCount, warn
         });
       }
     } catch (err) {
-      const range = `${batch[0].pageNum}-${batch[batch.length - 1].pageNum}`;
+      const range = `${batchPageNums[0]}-${batchPageNums[batchPageNums.length - 1]}`;
       warnings?.push({
         message: `\uBC30\uCE58 OCR \uC2E4\uD328 (\uD398\uC774\uC9C0 ${range}): ${err instanceof Error ? err.message : "\uC54C \uC218 \uC5C6\uB294 \uC624\uB958"}`,
         code: "OCR_PAGE_FAILED"
       });
-      for (const { pageNum } of batch) {
+      for (const pageNum of batchPageNums) {
         pageBlocks.push({ pageNum, blocks: [] });
       }
     }
-    processed += batch.length;
+    processed += batchPageNums.length;
     onProgress?.(processed, pageNumbers.length);
     return { batchIdx, pageBlocks };
   });
@@ -2752,24 +2759,29 @@ function isPdfFile(buffer) {
   const b = magicBytes(buffer);
   return b[0] === 37 && b[1] === 80 && b[2] === 68 && b[3] === 70;
 }
+function isPngFile(buffer) {
+  const b = magicBytes(buffer);
+  return b[0] === 137 && b[1] === 80 && b[2] === 78 && b[3] === 71;
+}
 function detectFormat(buffer) {
   if (buffer.byteLength < 4) return "unknown";
   if (isZipFile(buffer)) return "hwpx";
   if (isOldHwpFile(buffer)) return "hwp";
   if (isPdfFile(buffer)) return "pdf";
+  if (isPngFile(buffer)) return "image";
   return "unknown";
 }
 async function detectZipFormat(buffer) {
   try {
     const zip = await JSZip.loadAsync(buffer);
-    if (zip.file("xl/workbook.xml")) return "xlsx";
-    if (zip.file("word/document.xml")) return "docx";
-    if (zip.file("Contents/content.hpf") || zip.file("mimetype")) return "hwpx";
+    if (zip.file("xl/workbook.xml")) return { format: "xlsx", zip };
+    if (zip.file("word/document.xml")) return { format: "docx", zip };
+    if (zip.file("Contents/content.hpf") || zip.file("mimetype")) return { format: "hwpx", zip };
     const hasSection = Object.keys(zip.files).some((f) => f.startsWith("Contents/"));
-    if (hasSection) return "hwpx";
-    return "unknown";
+    if (hasSection) return { format: "hwpx", zip };
+    return { format: "unknown", zip: null };
   } catch {
-    return "unknown";
+    return { format: "unknown", zip: null };
   }
 }
 
@@ -2778,7 +2790,7 @@ import JSZip2 from "jszip";
 import { DOMParser } from "@xmldom/xmldom";
 
 // src/utils.ts
-var VERSION = true ? "2.3.1" : "0.0.0-dev";
+var VERSION = true ? "2.3.2" : "0.0.0-dev";
 function toArrayBuffer(buf) {
   if (buf.byteOffset === 0 && buf.byteLength === buf.buffer.byteLength) {
     return buf.buffer;
@@ -2938,12 +2950,16 @@ function buildTableDirect(rows, numRows) {
   return trimAndReturn(grid, numRows, maxCols);
 }
 function trimAndReturn(grid, numRows, maxCols) {
-  let effectiveCols = maxCols;
-  while (effectiveCols > 0) {
-    const colEmpty = grid.every((row) => !row[effectiveCols - 1]?.text?.trim());
-    if (!colEmpty) break;
-    effectiveCols--;
+  let effectiveCols = 0;
+  for (const row of grid) {
+    for (let c = row.length - 1; c >= effectiveCols; c--) {
+      if (row[c]?.text?.trim()) {
+        effectiveCols = c + 1;
+        break;
+      }
+    }
   }
+  if (effectiveCols === 0) effectiveCols = maxCols;
   if (effectiveCols < maxCols && effectiveCols > 0) {
     const trimmed = grid.map((row) => row.slice(0, effectiveCols));
     return { rows: numRows, cols: effectiveCols, cells: trimmed, hasHeader: numRows > 1 };
@@ -3200,11 +3216,11 @@ function parseStyleElements(doc, map) {
 function stripDtd(xml) {
   return xml.replace(/<!DOCTYPE\s[^[>]*(\[[\s\S]*?\])?\s*>/gi, "");
 }
-async function parseHwpxDocument(buffer, options) {
+async function parseHwpxDocument(buffer, options, existingZip) {
   precheckZipSize(buffer, MAX_DECOMPRESS_SIZE, MAX_ZIP_ENTRIES);
   let zip;
   try {
-    zip = await JSZip2.loadAsync(buffer);
+    zip = existingZip ?? await JSZip2.loadAsync(buffer);
   } catch {
     return await extractFromBrokenZip(buffer);
   }
@@ -6216,8 +6232,15 @@ import { getDocument, GlobalWorkerOptions } from "pdfjs-dist/legacy/build/pdf.mj
 GlobalWorkerOptions.workerSrc = "";
 var MAX_PAGES = 5e3;
 var MAX_TOTAL_TEXT = 500 * 1024 * 1024;
-var PDF_LOAD_TIMEOUT_MS = 3e4;
+function calcPdfTimeout(bufferSize) {
+  const base = 3e4;
+  const perMb = 500;
+  const mb = bufferSize / (1024 * 1024);
+  return Math.min(base + Math.ceil(mb * perMb), 3e5);
+}
 async function loadPdfWithTimeout(buffer) {
+  const timeoutMs = calcPdfTimeout(buffer.byteLength);
+  const timeoutSec = Math.round(timeoutMs / 1e3);
   const loadingTask = getDocument({
     data: new Uint8Array(buffer),
     useSystemFonts: true,
@@ -6231,8 +6254,8 @@ async function loadPdfWithTimeout(buffer) {
       new Promise((_, reject) => {
         timer = setTimeout(() => {
           loadingTask.destroy();
-          reject(new KordocError("PDF \uB85C\uB529 \uD0C0\uC784\uC544\uC6C3 (30\uCD08 \uCD08\uACFC)"));
-        }, PDF_LOAD_TIMEOUT_MS);
+          reject(new KordocError(`PDF \uB85C\uB529 \uD0C0\uC784\uC544\uC6C3 (${timeoutSec}\uCD08 \uCD08\uACFC)`));
+        }, timeoutMs);
       })
     ]);
   } finally {
@@ -6253,11 +6276,15 @@ async function parsePdfDocument(buffer, options) {
     const effectivePageCount = Math.min(pageCount, MAX_PAGES);
     const pageFilter = options?.pages ? parsePageRange(options.pages, effectivePageCount) : null;
     const totalTarget = pageFilter ? pageFilter.size : effectivePageCount;
-    const allFontSizes = [];
+    const fontSizeFreq = /* @__PURE__ */ new Map();
     const pageHeights = /* @__PURE__ */ new Map();
-    let parsedPages = 0;
+    const targetPageNums = [];
     for (let i = 1; i <= effectivePageCount; i++) {
       if (pageFilter && !pageFilter.has(i)) continue;
+      targetPageNums.push(i);
+    }
+    let parsedPages = 0;
+    const parseSinglePage = async (i) => {
       try {
         const page = await doc.getPage(i);
         const tc = await page.getTextContent();
@@ -6270,7 +6297,10 @@ async function parsePdfDocument(buffer, options) {
           warnings.push({ page: i, message: `${hiddenCount}\uAC1C \uC228\uACA8\uC9C4 \uD14D\uC2A4\uD2B8 \uC694\uC18C \uD544\uD130\uB9C1\uB428`, code: "HIDDEN_TEXT_FILTERED" });
         }
         for (const item of visible) {
-          if (item.fontSize > 0) allFontSizes.push(item.fontSize);
+          if (item.fontSize > 0) {
+            const rounded = Math.round(item.fontSize * 10) / 10;
+            fontSizeFreq.set(rounded, (fontSizeFreq.get(rounded) || 0) + 1);
+          }
         }
         const opList = await page.getOperatorList();
         const pageBlocks = extractPageBlocksWithLines(visible, i, opList, viewport.width, viewport.height);
@@ -6287,12 +6317,23 @@ async function parsePdfDocument(buffer, options) {
         if (pageErr instanceof KordocError) throw pageErr;
         warnings.push({ page: i, message: `\uD398\uC774\uC9C0 ${i} \uD30C\uC2F1 \uC2E4\uD328: ${pageErr instanceof Error ? pageErr.message : "\uC54C \uC218 \uC5C6\uB294 \uC624\uB958"}`, code: "PARTIAL_PARSE" });
       }
+    };
+    const sampleCount = Math.min(5, targetPageNums.length);
+    for (let si = 0; si < sampleCount; si++) {
+      await parseSinglePage(targetPageNums[si]);
+    }
+    const sampleParsed = parsedPages || sampleCount;
+    const isImageBased = totalChars / Math.max(sampleParsed, 1) < 10;
+    if (!isImageBased) {
+      for (let si = sampleCount; si < targetPageNums.length; si++) {
+        await parseSinglePage(targetPageNums[si]);
+      }
     }
     const parsedPageCount = parsedPages || (pageFilter ? pageFilter.size : effectivePageCount);
-    if (totalChars / Math.max(parsedPageCount, 1) < 10) {
+    if (isImageBased) {
       let ocrProvider = options?.ocr ?? null;
-      const ocrMode = options?.ocrMode;
-      if (!ocrProvider && ocrMode && ocrMode !== "off") {
+      const ocrMode = options?.ocrMode ?? "auto";
+      if (!ocrProvider && ocrMode !== "off") {
         try {
           const { resolveOcrProvider: resolveOcrProvider2 } = await Promise.resolve().then(() => (init_resolve(), resolve_exports));
           const concurrency = options?.ocrConcurrency ?? 1;
@@ -6344,7 +6385,7 @@ async function parsePdfDocument(buffer, options) {
         blocks.splice(removed[ri], 1);
       }
     }
-    const medianFontSize = computeMedianFontSize(allFontSizes);
+    const medianFontSize = computeMedianFromFreq(fontSizeFreq);
     if (medianFontSize > 0) {
       detectHeadings(blocks, medianFontSize);
     }
@@ -6397,11 +6438,18 @@ function filterHiddenText(items, pageWidth, pageHeight) {
   }
   return { visible, hiddenCount };
 }
-function computeMedianFontSize(sizes) {
-  if (sizes.length === 0) return 0;
-  const sorted = [...sizes].sort((a, b) => a - b);
-  const mid = Math.floor(sorted.length / 2);
-  return sorted.length % 2 === 0 ? (sorted[mid - 1] + sorted[mid]) / 2 : sorted[mid];
+function computeMedianFromFreq(freq) {
+  if (freq.size === 0) return 0;
+  const entries = [...freq.entries()].sort((a, b) => a[0] - b[0]);
+  let total = 0;
+  for (const [, count] of entries) total += count;
+  const mid = total / 2;
+  let cumulative = 0;
+  for (const [size, count] of entries) {
+    cumulative += count;
+    if (cumulative >= mid) return size;
+  }
+  return 0;
 }
 function detectHeadings(blocks, medianFontSize) {
   for (const block of blocks) {
@@ -7204,6 +7252,7 @@ var MAX_SHEETS = 100;
 var MAX_DECOMPRESS_SIZE3 = 500 * 1024 * 1024;
 var MAX_ROWS2 = 1e4;
 var MAX_COLS2 = 200;
+var MAX_TOTAL_CELLS = 2e6;
 function cleanNumericValue(raw) {
   if (!/^-?\d+\.\d+$/.test(raw)) return raw;
   const num = parseFloat(raw);
@@ -7387,9 +7436,9 @@ function sheetToBlocks(sheetName, grid, merges, maxRow, maxCol, sheetIndex) {
   }
   return blocks;
 }
-async function parseXlsxDocument(buffer, options) {
+async function parseXlsxDocument(buffer, options, existingZip) {
   precheckZipSize(buffer, MAX_DECOMPRESS_SIZE3);
-  const zip = await JSZip3.loadAsync(buffer);
+  const zip = existingZip ?? await JSZip3.loadAsync(buffer);
   const warnings = [];
   const workbookFile = zip.file("xl/workbook.xml");
   if (!workbookFile) {
@@ -7416,6 +7465,7 @@ async function parseXlsxDocument(buffer, options) {
   }
   const blocks = [];
   const processedSheets = Math.min(sheets.length, MAX_SHEETS);
+  let totalCells = 0;
   for (let i = 0; i < processedSheets; i++) {
     if (pageFilter && !pageFilter.has(i + 1)) continue;
     const sheet = sheets[i];
@@ -7442,6 +7492,11 @@ async function parseXlsxDocument(buffer, options) {
     try {
       const sheetXml = await sheetFile.async("text");
       const { grid, merges, maxRow, maxCol } = parseWorksheet(sheetXml, sharedStrings);
+      totalCells += maxRow * maxCol;
+      if (totalCells > MAX_TOTAL_CELLS) {
+        warnings.push({ message: `\uCD1D \uC140 \uC218 \uC81C\uD55C \uCD08\uACFC (${totalCells.toLocaleString()}\uC140), \uC774\uD6C4 \uC2DC\uD2B8 \uC0DD\uB7B5`, code: "PARTIAL_PARSE" });
+        break;
+      }
       const sheetBlocks = sheetToBlocks(sheet.name, grid, merges, maxRow, maxCol, i);
       blocks.push(...sheetBlocks);
     } catch (err) {
@@ -7525,10 +7580,35 @@ function getAttr(el, localName) {
 function parseXml2(text) {
   return new DOMParser3().parseFromString(text, "text/xml");
 }
+function buildElementIndex(root) {
+  const index = /* @__PURE__ */ new Map();
+  const walk = (node) => {
+    const children = node.childNodes;
+    for (let i = 0; i < children.length; i++) {
+      const child = children[i];
+      if (child.nodeType === 1) {
+        const el = child;
+        const name = el.localName ?? "";
+        if (name) {
+          let list = index.get(name);
+          if (!list) {
+            list = [];
+            index.set(name, list);
+          }
+          list.push(el);
+        }
+        walk(el);
+      }
+    }
+  };
+  walk(root);
+  return index;
+}
 function parseStyles(xml) {
   const doc = parseXml2(xml);
   const styles = /* @__PURE__ */ new Map();
-  const styleElements = findElements(doc, "style");
+  const idx = buildElementIndex(doc);
+  const styleElements = idx.get("style") ?? [];
   for (const el of styleElements) {
     const styleId = getAttr(el, "styleId");
     if (!styleId) continue;
@@ -7556,7 +7636,8 @@ function parseStyles(xml) {
 function parseNumbering(xml) {
   const doc = parseXml2(xml);
   const abstractNums = /* @__PURE__ */ new Map();
-  const abstractElements = findElements(doc, "abstractNum");
+  const idx = buildElementIndex(doc);
+  const abstractElements = idx.get("abstractNum") ?? [];
   for (const el of abstractElements) {
     const abstractNumId = getAttr(el, "abstractNumId");
     if (!abstractNumId) continue;
@@ -7571,7 +7652,7 @@ function parseNumbering(xml) {
     abstractNums.set(abstractNumId, levels);
   }
   const nums = /* @__PURE__ */ new Map();
-  const numElements = findElements(doc, "num");
+  const numElements = idx.get("num") ?? [];
   for (const el of numElements) {
     const numId = getAttr(el, "numId");
     if (!numId) continue;
@@ -7815,9 +7896,9 @@ async function extractImages(zip, rels, doc) {
   }
   return { blocks, images };
 }
-async function parseDocxDocument(buffer, options) {
+async function parseDocxDocument(buffer, options, existingZip) {
   precheckZipSize(buffer, MAX_DECOMPRESS_SIZE4);
-  const zip = await JSZip4.loadAsync(buffer);
+  const zip = existingZip ?? await JSZip4.loadAsync(buffer);
   const warnings = [];
   const docFile = zip.file("word/document.xml");
   if (!docFile) {
@@ -7907,6 +7988,11 @@ async function parseDocxDocument(buffer, options) {
   };
 }
 
+// src/index.ts
+init_cli_provider();
+init_tesseract_provider();
+init_markdown_to_blocks();
+
 // src/diff/text-diff.ts
 function similarity(a, b) {
   if (a === b) return 1;
@@ -10423,25 +10509,86 @@ async function parse2(input, options) {
   if (!buffer || buffer.byteLength === 0) {
     return { success: false, fileType: "unknown", error: "\uBE48 \uBC84\uD37C\uC774\uAC70\uB098 \uC720\uD6A8\uD558\uC9C0 \uC54A\uC740 \uC785\uB825\uC785\uB2C8\uB2E4.", code: "EMPTY_INPUT" };
   }
+  const MAX_FILE_SIZE = 500 * 1024 * 1024;
+  if (buffer.byteLength > MAX_FILE_SIZE) {
+    return { success: false, fileType: "unknown", error: `\uD30C\uC77C \uD06C\uAE30 \uCD08\uACFC: ${(buffer.byteLength / 1024 / 1024).toFixed(1)}MB (\uCD5C\uB300 500MB)`, code: "FILE_TOO_LARGE" };
+  }
   const format = detectFormat(buffer);
   switch (format) {
     case "hwpx": {
-      const zipFormat = await detectZipFormat(buffer);
-      if (zipFormat === "xlsx") return parseXlsx(buffer, options);
-      if (zipFormat === "docx") return parseDocx(buffer, options);
-      return parseHwpx(buffer, options);
+      const { format: zipFormat, zip } = await detectZipFormat(buffer);
+      if (zipFormat === "xlsx") return parseXlsx(buffer, options, zip ?? void 0);
+      if (zipFormat === "docx") return parseDocx(buffer, options, zip ?? void 0);
+      return parseHwpx(buffer, options, zip ?? void 0);
     }
     case "hwp":
       return parseHwp(buffer, options);
     case "pdf":
       return parsePdf(buffer, options);
+    case "image":
+      return parseImage(buffer, options);
     default:
       return { success: false, fileType: "unknown", error: "\uC9C0\uC6D0\uD558\uC9C0 \uC54A\uB294 \uD30C\uC77C \uD615\uC2DD\uC785\uB2C8\uB2E4.", code: "UNSUPPORTED_FORMAT" };
   }
 }
-async function parseHwpx(buffer, options) {
+async function parseImage(buffer, options) {
+  const ocrMode = options?.ocrMode || "auto";
+  if (ocrMode === "off") {
+    return { success: false, fileType: "image", error: "OCR\uC774 \uBE44\uD65C\uC131\uD654\uB418\uC5B4 \uC774\uBBF8\uC9C0 \uD30C\uC77C\uC744 \uCC98\uB9AC\uD560 \uC218 \uC5C6\uC2B5\uB2C8\uB2E4.", code: "IMAGE_BASED_PDF" };
+  }
+  let ocrProvider;
+  let actualOcrMode = "auto";
+  try {
+    if (ocrMode === "gemini" || ocrMode === "claude" || ocrMode === "codex" || ocrMode === "ollama") {
+      ocrProvider = createCliOcrProvider(ocrMode);
+      actualOcrMode = ocrMode;
+    } else if (ocrMode === "tesseract") {
+      ocrProvider = await createTesseractProvider();
+      actualOcrMode = ocrMode;
+    } else if (ocrMode === "auto") {
+      const modesToTry = ["gemini", "claude", "codex", "ollama"];
+      for (const mode of modesToTry) {
+        try {
+          ocrProvider = createCliOcrProvider(mode);
+          actualOcrMode = mode;
+          break;
+        } catch (e) {
+          console.warn(`[kordoc] OCR auto-detection: ${mode} CLI not available or failed. Trying next.`, e);
+        }
+      }
+      if (!ocrProvider) {
+        ocrProvider = await createTesseractProvider();
+        actualOcrMode = "tesseract";
+      }
+    }
+    if (!ocrProvider) {
+      return { success: false, fileType: "image", error: "\uC0AC\uC6A9 \uAC00\uB2A5\uD55C OCR \uD504\uB85C\uBC14\uC774\uB354\uB97C \uCC3E\uC744 \uC218 \uC5C6\uC2B5\uB2C8\uB2E4.", code: "PARSE_ERROR" };
+    }
+    const imageUint8Array = new Uint8Array(buffer);
+    const ocrResult = await ocrProvider(imageUint8Array, 1, "image/png");
+    if (ocrProvider.terminate) {
+      await ocrProvider.terminate();
+    }
+    const markdown = typeof ocrResult === "string" ? ocrResult : ocrResult.markdown;
+    const blocks = markdownToBlocks(markdown, 1);
+    return {
+      success: true,
+      fileType: "image",
+      markdown,
+      blocks,
+      isImageBased: true,
+      warnings: [{ message: `OCR \uCC98\uB9AC\uB428 (${actualOcrMode})`, code: "OCR_FALLBACK" }]
+    };
+  } catch (err) {
+    if (ocrProvider && ocrProvider.terminate) {
+      await ocrProvider.terminate();
+    }
+    return { success: false, fileType: "image", error: err instanceof Error ? err.message : "\uC774\uBBF8\uC9C0 OCR \uC2E4\uD328", code: classifyError(err) };
+  }
+}
+async function parseHwpx(buffer, options, zip) {
   try {
-    const { markdown, blocks, metadata, outline, warnings, images } = await parseHwpxDocument(buffer, options);
+    const { markdown, blocks, metadata, outline, warnings, images } = await parseHwpxDocument(buffer, options, zip);
     return { success: true, fileType: "hwpx", markdown, blocks, metadata, outline, warnings, images: images?.length ? images : void 0 };
   } catch (err) {
     return { success: false, fileType: "hwpx", error: err instanceof Error ? err.message : "HWPX \uD30C\uC2F1 \uC2E4\uD328", code: classifyError(err) };
@@ -10464,17 +10611,17 @@ async function parsePdf(buffer, options) {
     return { success: false, fileType: "pdf", error: err instanceof Error ? err.message : "PDF \uD30C\uC2F1 \uC2E4\uD328", code: classifyError(err), isImageBased };
   }
 }
-async function parseXlsx(buffer, options) {
+async function parseXlsx(buffer, options, zip) {
   try {
-    const { markdown, blocks, metadata, warnings } = await parseXlsxDocument(buffer, options);
+    const { markdown, blocks, metadata, warnings } = await parseXlsxDocument(buffer, options, zip);
     return { success: true, fileType: "xlsx", markdown, blocks, metadata, warnings };
   } catch (err) {
     return { success: false, fileType: "xlsx", error: err instanceof Error ? err.message : "XLSX \uD30C\uC2F1 \uC2E4\uD328", code: classifyError(err) };
   }
 }
-async function parseDocx(buffer, options) {
+async function parseDocx(buffer, options, zip) {
   try {
-    const { markdown, blocks, metadata, outline, warnings, images } = await parseDocxDocument(buffer, options);
+    const { markdown, blocks, metadata, outline, warnings, images } = await parseDocxDocument(buffer, options, zip);
     return { success: true, fileType: "docx", markdown, blocks, metadata, outline, warnings, images: images?.length ? images : void 0 };
   } catch (err) {
     return { success: false, fileType: "docx", error: err instanceof Error ? err.message : "DOCX \uD30C\uC2F1 \uC2E4\uD328", code: classifyError(err) };