@clazic/kordoc 2.3.1 → 2.3.3

package/dist/{batch-provider-PCT4I4LK.js → batch-provider-PNDCSGQW.js}

@@ -108,9 +108,8 @@ 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;
@@ -166,4 +165,4 @@ export {
   DEFAULT_BATCH_SIZES,
   createBatchCliProvider
 };
-//# sourceMappingURL=batch-provider-PCT4I4LK.js.map
+//# sourceMappingURL=batch-provider-PNDCSGQW.js.map

package/dist/batch-provider-PNDCSGQW.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/ocr/batch-provider.ts"],"sourcesContent":["/**\n * CLI 배치 OCR 프로바이더\n *\n * 여러 페이지 이미지를 단일 CLI 호출로 처리하여 API 호출 수를 대폭 감소.\n * gemini/claude: @file 멀티 참조, codex: --image 멀티 플래그\n *\n * 299페이지 기준:\n * - 기존: CLI 299회 호출 (~30분)\n * - 배치: CLI 3~6회 호출 (~3분)\n */\n\nimport { spawn } from \"child_process\"\nimport { writeFileSync, readFileSync, unlinkSync, mkdirSync } from \"fs\"\nimport { join } from \"path\"\nimport { tmpdir } from \"os\"\nimport type { StructuredOcrResult, BatchOcrProvider } from \"../types.js\"\n\n/** 배치 OCR 프롬프트 */\nconst BATCH_OCR_PROMPT =\n  \"다음 문서 페이지 이미지들을 OCR하여 순수 Markdown으로 변환하세요.\\n\\n\" +\n  \"규칙:\\n\" +\n  \"- 각 페이지 결과 사이에 반드시 이 구분자를 삽입: <!-- PAGE_BREAK -->\\n\" +\n  \"- 테이블은 Markdown 테이블 문법 사용 (| 구분, |---|---| 헤더 구분선 포함)\\n\" +\n  \"- 병합된 셀은 해당 위치에 내용 기재\\n\" +\n  \"- 헤딩은 글자 크기에 따라 ## ~ ###### 사용\\n\" +\n  \"- 리스트는 - 또는 1. 사용\\n\" +\n  \"- 이미지, 도형 등 비텍스트 요소는 무시\\n\" +\n  \"- 원문의 읽기 순서와 구조를 유지\\n\" +\n  \"- ```로 감싸지 말고 순수 Markdown만 출력\"\n\n/** 모드별 기본 배치 크기 (CLI 내부 타임아웃 + 실측 기반)\n *\n * gemini CLI: 10장 이상에서 AbortError 발생 (내부 타임아웃).\n * 5장 배치가 안정적으로 동작 확인 (35초/배치).\n * 299페이지 = 60배치 = 기존 299회 대비 80% 감소.\n */\nexport const DEFAULT_BATCH_SIZES: Record<string, number> = {\n  gemini: 5,\n  claude: 5,\n  codex: 10,\n}\n\n/** 임시 디렉토리 — gemini CLI는 cwd 하위 + gitignore 밖만 @참조 가능 */\nlet _batchTempDir: string | null = null\nfunction getBatchTempDir(): string {\n  if (!_batchTempDir) {\n    _batchTempDir = join(process.cwd(), \"_kordoc_ocr_tmp\")\n    mkdirSync(_batchTempDir, { recursive: true })\n  }\n  return _batchTempDir\n}\n\n/**\n * 배치 CLI 프로바이더 생성\n */\nexport function createBatchCliProvider(\n  mode: \"gemini\" | \"claude\" | \"codex\",\n  batchSize: number\n): BatchOcrProvider {\n  return {\n    __batch: true as const,\n    batchSize,\n    async processBatch(pages) {\n      const results = new Map<number, StructuredOcrResult>()\n      const tempDir = getBatchTempDir()\n      const tempFiles: string[] = []\n\n      try {\n        // 1. Write all page images to temp files\n        for (const { image, pageNum } of pages) {\n          const path = join(tempDir, `batch-p${pageNum}.png`)\n          writeFileSync(path, image)\n          tempFiles.push(path)\n        }\n\n        // 2. Call CLI with all file references (비동기 — 병렬 배치 실행 가능)\n        let output: string\n        if (mode === \"codex\") {\n          output = await callBatchCodexCli(tempFiles)\n        } else {\n          output = await callBatchCli(mode, tempFiles)\n        }\n\n        // 3. Parse response by PAGE_BREAK separator\n        const cleaned = stripCodeFence(output.trim())\n        const parts = cleaned.split(/<!--\\s*PAGE_BREAK\\s*-->/)\n          .map(p => p.trim())\n          .filter(p => p.length > 0)\n\n        // 4. Map results to page numbers (best-effort if count mismatch)\n        for (let i = 0; i < pages.length; i++) {\n          const pageNum = pages[i].pageNum\n          if (i < parts.length) {\n            results.set(pageNum, { markdown: parts[i] })\n          }\n          // If fewer parts than pages, remaining pages get no result\n        }\n      } finally {\n        // 5. Clean up temp files\n        for (const f of tempFiles) {\n          try { unlinkSync(f) } catch { /* ignore */ }\n        }\n      }\n\n      return results\n    },\n  }\n}\n\n/**\n * 비동기 CLI 실행 헬퍼 — spawn + Promise 래핑.\n * spawnSync는 이벤트 루프를 차단하여 병렬 배치 실행 불가.\n */\nfunction spawnAsync(\n  cmd: string,\n  args: string[],\n  opts: { timeoutMs: number; cwd?: string; stdin?: string }\n): Promise<{ stdout: string; stderr: string; exitCode: number }> {\n  return new Promise((resolve, reject) => {\n    const child = spawn(cmd, args, {\n      cwd: opts.cwd,\n      env: process.env,\n      stdio: [\"pipe\", \"pipe\", \"pipe\"],\n    })\n\n    let stdout = \"\"\n    let stderr = \"\"\n    let killed = false\n\n    child.stdout.setEncoding(\"utf-8\")\n    child.stderr.setEncoding(\"utf-8\")\n    child.stdout.on(\"data\", (d: string) => { stdout += d })\n    child.stderr.on(\"data\", (d: string) => { stderr += d })\n\n    const timer = setTimeout(() => {\n      killed = true\n      child.kill(\"SIGTERM\")\n    }, opts.timeoutMs)\n\n    if (opts.stdin !== undefined) {\n      child.stdin.end(opts.stdin)\n    } else {\n      child.stdin.end()\n    }\n\n    child.on(\"close\", (code) => {\n      clearTimeout(timer)\n      if (killed) {\n        reject(new Error(`타임아웃 (${Math.round(opts.timeoutMs / 1000)}초)`))\n      } else {\n        resolve({ stdout, stderr, exitCode: code ?? 1 })\n      }\n    })\n    child.on(\"error\", (err) => {\n      clearTimeout(timer)\n      reject(err)\n    })\n  })\n}\n\n/** gemini/claude 배치 호출 (비동기) */\nasync function callBatchCli(mode: \"gemini\" | \"claude\", imagePaths: string[]): Promise<string> {\n  const fileRefs = imagePaths.map(p => `@${p}`).join(\"\\n\")\n  const prompt = `${BATCH_OCR_PROMPT}\\n\\n${fileRefs}`\n\n  let args: string[]\n  if (mode === \"gemini\") {\n    const model = process.env.KORDOC_GEMINI_MODEL ?? \"gemini-2.5-flash\"\n    args = [\"--prompt\", prompt, \"--yolo\", \"--model\", model]\n  } else {\n    args = [\"--print\", prompt]\n    const model = process.env.KORDOC_CLAUDE_MODEL\n    if (model) args.push(\"--model\", model)\n  }\n\n  const timeoutMs = 60_000 + imagePaths.length * 20_000\n  const result = await spawnAsync(mode, args, {\n    timeoutMs,\n    ...(mode === \"claude\" ? { cwd: tmpdir() } : {}),\n  })\n\n  if (result.exitCode !== 0) {\n    const errMsg = result.stderr?.trim() || `exit code ${result.exitCode}`\n    throw new Error(`${mode} 배치 OCR 실패: ${errMsg}`)\n  }\n\n  return result.stdout || \"\"\n}\n\n/** codex 배치 호출 (비동기) — --image를 여러 번 지정 */\nasync function callBatchCodexCli(imagePaths: string[]): Promise<string> {\n  const outPath = join(tmpdir(), `kordoc-codex-batch-${Date.now()}-${Math.random().toString(36).slice(2)}.txt`)\n  try {\n    const args = [\"exec\", BATCH_OCR_PROMPT]\n    for (const p of imagePaths) {\n      args.push(\"--image\", p)\n    }\n    args.push(\"--output-last-message\", outPath)\n    const model = process.env.KORDOC_CODEX_MODEL\n    if (model) args.push(\"--model\", model)\n\n    const timeoutMs = 60_000 + imagePaths.length * 20_000\n    const result = await spawnAsync(\"codex\", args, {\n      timeoutMs,\n      stdin: \"\",\n    })\n\n    if (result.exitCode !== 0) {\n      const errMsg = result.stderr?.trim() || `exit code ${result.exitCode}`\n      throw new Error(`codex 배치 OCR 실패: ${errMsg}`)\n    }\n\n    try {\n      return readFileSync(outPath, \"utf-8\")\n    } catch {\n      return result.stdout || \"\"\n    }\n  } finally {\n    try { unlinkSync(outPath) } catch { /* ignore */ }\n  }\n}\n\n/** LLM 출력에서 코드 펜스 제거 (cli-provider.ts와 동일 로직) */\nfunction stripCodeFence(text: string): string {\n  const match = text.match(/^```(?:markdown|md)?\\s*\\n([\\s\\S]*?)\\n```\\s*$/m)\n  return match ? match[1].trim() : text\n}\n"],"mappings":";;;;AAWA,SAAS,aAAa;AACtB,SAAS,eAAe,cAAc,YAAY,iBAAiB;AACnE,SAAS,YAAY;AACrB,SAAS,cAAc;AAIvB,IAAM,mBACJ;AAiBK,IAAM,sBAA8C;AAAA,EACzD,QAAQ;AAAA,EACR,QAAQ;AAAA,EACR,OAAO;AACT;AAGA,IAAI,gBAA+B;AACnC,SAAS,kBAA0B;AACjC,MAAI,CAAC,eAAe;AAClB,oBAAgB,KAAK,QAAQ,IAAI,GAAG,iBAAiB;AACrD,cAAU,eAAe,EAAE,WAAW,KAAK,CAAC;AAAA,EAC9C;AACA,SAAO;AACT;AAKO,SAAS,uBACd,MACA,WACkB;AAClB,SAAO;AAAA,IACL,SAAS;AAAA,IACT;AAAA,IACA,MAAM,aAAa,OAAO;AACxB,YAAM,UAAU,oBAAI,IAAiC;AACrD,YAAM,UAAU,gBAAgB;AAChC,YAAM,YAAsB,CAAC;AAE7B,UAAI;AAEF,mBAAW,EAAE,OAAO,QAAQ,KAAK,OAAO;AACtC,gBAAM,OAAO,KAAK,SAAS,UAAU,OAAO,MAAM;AAClD,wBAAc,MAAM,KAAK;AACzB,oBAAU,KAAK,IAAI;AAAA,QACrB;AAGA,YAAI;AACJ,YAAI,SAAS,SAAS;AACpB,mBAAS,MAAM,kBAAkB,SAAS;AAAA,QAC5C,OAAO;AACL,mBAAS,MAAM,aAAa,MAAM,SAAS;AAAA,QAC7C;AAGA,cAAM,UAAU,eAAe,OAAO,KAAK,CAAC;AAC5C,cAAM,QAAQ,QAAQ,MAAM,yBAAyB,EAClD,IAAI,OAAK,EAAE,KAAK,CAAC,EACjB,OAAO,OAAK,EAAE,SAAS,CAAC;AAG3B,iBAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,gBAAM,UAAU,MAAM,CAAC,EAAE;AACzB,cAAI,IAAI,MAAM,QAAQ;AACpB,oBAAQ,IAAI,SAAS,EAAE,UAAU,MAAM,CAAC,EAAE,CAAC;AAAA,UAC7C;AAAA,QAEF;AAAA,MACF,UAAE;AAEA,mBAAW,KAAK,WAAW;AACzB,cAAI;AAAE,uBAAW,CAAC;AAAA,UAAE,QAAQ;AAAA,UAAe;AAAA,QAC7C;AAAA,MACF;AAEA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAMA,SAAS,WACP,KACA,MACA,MAC+D;AAC/D,SAAO,IAAI,QAAQ,CAAC,SAAS,WAAW;AACtC,UAAM,QAAQ,MAAM,KAAK,MAAM;AAAA,MAC7B,KAAK,KAAK;AAAA,MACV,KAAK,QAAQ;AAAA,MACb,OAAO,CAAC,QAAQ,QAAQ,MAAM;AAAA,IAChC,CAAC;AAED,QAAI,SAAS;AACb,QAAI,SAAS;AACb,QAAI,SAAS;AAEb,UAAM,OAAO,YAAY,OAAO;AAChC,UAAM,OAAO,YAAY,OAAO;AAChC,UAAM,OAAO,GAAG,QAAQ,CAAC,MAAc;AAAE,gBAAU;AAAA,IAAE,CAAC;AACtD,UAAM,OAAO,GAAG,QAAQ,CAAC,MAAc;AAAE,gBAAU;AAAA,IAAE,CAAC;AAEtD,UAAM,QAAQ,WAAW,MAAM;AAC7B,eAAS;AACT,YAAM,KAAK,SAAS;AAAA,IACtB,GAAG,KAAK,SAAS;AAEjB,QAAI,KAAK,UAAU,QAAW;AAC5B,YAAM,MAAM,IAAI,KAAK,KAAK;AAAA,IAC5B,OAAO;AACL,YAAM,MAAM,IAAI;AAAA,IAClB;AAEA,UAAM,GAAG,SAAS,CAAC,SAAS;AAC1B,mBAAa,KAAK;AAClB,UAAI,QAAQ;AACV,eAAO,IAAI,MAAM,6BAAS,KAAK,MAAM,KAAK,YAAY,GAAI,CAAC,SAAI,CAAC;AAAA,MAClE,OAAO;AACL,gBAAQ,EAAE,QAAQ,QAAQ,UAAU,QAAQ,EAAE,CAAC;AAAA,MACjD;AAAA,IACF,CAAC;AACD,UAAM,GAAG,SAAS,CAAC,QAAQ;AACzB,mBAAa,KAAK;AAClB,aAAO,GAAG;AAAA,IACZ,CAAC;AAAA,EACH,CAAC;AACH;AAGA,eAAe,aAAa,MAA2B,YAAuC;AAC5F,QAAM,WAAW,WAAW,IAAI,OAAK,IAAI,CAAC,EAAE,EAAE,KAAK,IAAI;AACvD,QAAM,SAAS,GAAG,gBAAgB;AAAA;AAAA,EAAO,QAAQ;AAEjD,MAAI;AACJ,MAAI,SAAS,UAAU;AACrB,UAAM,QAAQ,QAAQ,IAAI,uBAAuB;AACjD,WAAO,CAAC,YAAY,QAAQ,UAAU,WAAW,KAAK;AAAA,EACxD,OAAO;AACL,WAAO,CAAC,WAAW,MAAM;AACzB,UAAM,QAAQ,QAAQ,IAAI;AAC1B,QAAI,MAAO,MAAK,KAAK,WAAW,KAAK;AAAA,EACvC;AAEA,QAAM,YAAY,MAAS,WAAW,SAAS;AAC/C,QAAM,SAAS,MAAM,WAAW,MAAM,MAAM;AAAA,IAC1C;AAAA,IACA,GAAI,SAAS,WAAW,EAAE,KAAK,OAAO,EAAE,IAAI,CAAC;AAAA,EAC/C,CAAC;AAED,MAAI,OAAO,aAAa,GAAG;AACzB,UAAM,SAAS,OAAO,QAAQ,KAAK,KAAK,aAAa,OAAO,QAAQ;AACpE,UAAM,IAAI,MAAM,GAAG,IAAI,mCAAe,MAAM,EAAE;AAAA,EAChD;AAEA,SAAO,OAAO,UAAU;AAC1B;AAGA,eAAe,kBAAkB,YAAuC;AACtE,QAAM,UAAU,KAAK,OAAO,GAAG,sBAAsB,KAAK,IAAI,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,CAAC,CAAC,MAAM;AAC5G,MAAI;AACF,UAAM,OAAO,CAAC,QAAQ,gBAAgB;AACtC,eAAW,KAAK,YAAY;AAC1B,WAAK,KAAK,WAAW,CAAC;AAAA,IACxB;AACA,SAAK,KAAK,yBAAyB,OAAO;AAC1C,UAAM,QAAQ,QAAQ,IAAI;AAC1B,QAAI,MAAO,MAAK,KAAK,WAAW,KAAK;AAErC,UAAM,YAAY,MAAS,WAAW,SAAS;AAC/C,UAAM,SAAS,MAAM,WAAW,SAAS,MAAM;AAAA,MAC7C;AAAA,MACA,OAAO;AAAA,IACT,CAAC;AAED,QAAI,OAAO,aAAa,GAAG;AACzB,YAAM,SAAS,OAAO,QAAQ,KAAK,KAAK,aAAa,OAAO,QAAQ;AACpE,YAAM,IAAI,MAAM,wCAAoB,MAAM,EAAE;AAAA,IAC9C;AAEA,QAAI;AACF,aAAO,aAAa,SAAS,OAAO;AAAA,IACtC,QAAQ;AACN,aAAO,OAAO,UAAU;AAAA,IAC1B;AAAA,EACF,UAAE;AACA,QAAI;AAAE,iBAAW,OAAO;AAAA,IAAE,QAAQ;AAAA,IAAe;AAAA,EACnD;AACF;AAGA,SAAS,eAAe,MAAsB;AAC5C,QAAM,QAAQ,KAAK,MAAM,+CAA+C;AACxE,SAAO,QAAQ,MAAM,CAAC,EAAE,KAAK,IAAI;AACnC;","names":[]}

package/dist/chunk-4PP34NVQ.js

@@ -0,0 +1,121 @@
+#!/usr/bin/env node
+
+// src/ocr/markdown-to-blocks.ts
+function markdownToBlocks(markdown, pageNumber) {
+  const blocks = [];
+  const lines = markdown.split("\n");
+  let i = 0;
+  while (i < lines.length) {
+    const line = lines[i];
+    if (line.trim() === "") {
+      i++;
+      continue;
+    }
+    const headingMatch = line.match(/^(#{1,6})\s+(.+)$/);
+    if (headingMatch) {
+      blocks.push({
+        type: "heading",
+        level: headingMatch[1].length,
+        text: headingMatch[2].trim(),
+        pageNumber
+      });
+      i++;
+      continue;
+    }
+    if (/^[-*_]{3,}\s*$/.test(line.trim())) {
+      blocks.push({ type: "separator", pageNumber });
+      i++;
+      continue;
+    }
+    if (line.trim().startsWith("|")) {
+      const tableLines = [];
+      while (i < lines.length && lines[i].trim().startsWith("|")) {
+        tableLines.push(lines[i]);
+        i++;
+      }
+      const table = parseMarkdownTable(tableLines);
+      if (table) {
+        blocks.push({ type: "table", table, pageNumber });
+      }
+      continue;
+    }
+    const ulMatch = line.match(/^(\s*)[-*+]\s+(.+)$/);
+    if (ulMatch) {
+      blocks.push({
+        type: "list",
+        listType: "unordered",
+        text: ulMatch[2].trim(),
+        pageNumber
+      });
+      i++;
+      continue;
+    }
+    const olMatch = line.match(/^(\s*)\d+\.\s+(.+)$/);
+    if (olMatch) {
+      blocks.push({
+        type: "list",
+        listType: "ordered",
+        text: olMatch[2].trim(),
+        pageNumber
+      });
+      i++;
+      continue;
+    }
+    const paraLines = [];
+    while (i < lines.length && lines[i].trim() !== "" && !isStructuralLine(lines[i])) {
+      paraLines.push(lines[i].trim());
+      i++;
+    }
+    if (paraLines.length > 0) {
+      blocks.push({
+        type: "paragraph",
+        text: paraLines.join("\n"),
+        pageNumber
+      });
+    }
+  }
+  return blocks;
+}
+function isStructuralLine(line) {
+  if (/^#{1,6}\s+/.test(line)) return true;
+  if (line.trim().startsWith("|")) return true;
+  if (/^[-*_]{3,}\s*$/.test(line.trim())) return true;
+  if (/^\s*[-*+]\s+/.test(line)) return true;
+  if (/^\s*\d+\.\s+/.test(line)) return true;
+  return false;
+}
+function parseMarkdownTable(lines) {
+  const hasSeparator = lines.some((line) => /^\|[\s:|-]+\|$/.test(line.trim()));
+  const rows = [];
+  let maxCols = 0;
+  for (const line of lines) {
+    if (/^\|\s*:?-+:?\s*(\|\s*:?-+:?\s*)+\|?\s*$/.test(line.trim())) continue;
+    const parts = line.split("|");
+    const cells = parts.slice(1, parts[parts.length - 1].trim() === "" ? -1 : void 0).map((cell) => ({
+      text: cell.trim(),
+      colSpan: 1,
+      rowSpan: 1
+    }));
+    if (cells.length > 0) {
+      rows.push(cells);
+      maxCols = Math.max(maxCols, cells.length);
+    }
+  }
+  if (rows.length === 0) return null;
+  for (const row of rows) {
+    while (row.length < maxCols) {
+      row.push({ text: "", colSpan: 1, rowSpan: 1 });
+    }
+  }
+  return {
+    rows: rows.length,
+    cols: maxCols,
+    cells: rows,
+    hasHeader: hasSeparator && rows.length > 1
+  };
+}
+
+export {
+  markdownToBlocks
+};
+//# sourceMappingURL=chunk-4PP34NVQ.js.map

package/dist/chunk-4PP34NVQ.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/ocr/markdown-to-blocks.ts"],"sourcesContent":["/**\n * Markdown → IRBlock[] 역파싱\n *\n * Vision LLM(gemini/claude/codex 등)이 반환한 Markdown 문자열을\n * kordoc의 IRBlock[] 중간 표현으로 변환.\n * 기존 blocksToMarkdown()의 역방향 처리.\n */\n\nimport type { IRBlock, IRTable, IRCell } from \"../types.js\"\n\n/**\n * Markdown 문자열을 IRBlock[] 배열로 변환.\n *\n * 지원 요소:\n * - 헤딩: # ~ ######\n * - 테이블: | col1 | col2 | (파이프 구분, |---|---| 구분선 포함)\n * - 순서/비순서 리스트: - / 1.\n * - 구분선: ---, ***, ___\n * - 일반 텍스트 (paragraph)\n */\nexport function markdownToBlocks(markdown: string, pageNumber: number): IRBlock[] {\n  const blocks: IRBlock[] = []\n  const lines = markdown.split(\"\\n\")\n  let i = 0\n\n  while (i < lines.length) {\n    const line = lines[i]\n\n    // 빈 줄 스킵\n    if (line.trim() === \"\") {\n      i++\n      continue\n    }\n\n    // 1. 헤딩: # ~ ######\n    const headingMatch = line.match(/^(#{1,6})\\s+(.+)$/)\n    if (headingMatch) {\n      blocks.push({\n        type: \"heading\",\n        level: headingMatch[1].length,\n        text: headingMatch[2].trim(),\n        pageNumber,\n      })\n      i++\n      continue\n    }\n\n    // 2. 구분선: ---, ***, ___\n    if (/^[-*_]{3,}\\s*$/.test(line.trim())) {\n      blocks.push({ type: \"separator\", pageNumber })\n      i++\n      continue\n    }\n\n    // 3. 테이블: | 로 시작하는 연속 행 수집\n    if (line.trim().startsWith(\"|\")) {\n      const tableLines: string[] = []\n      while (i < lines.length && lines[i].trim().startsWith(\"|\")) {\n        tableLines.push(lines[i])\n        i++\n      }\n      const table = parseMarkdownTable(tableLines)\n      if (table) {\n        blocks.push({ type: \"table\", table, pageNumber })\n      }\n      continue\n    }\n\n    // 4. 비순서 리스트: -, *, +\n    const ulMatch = line.match(/^(\\s*)[-*+]\\s+(.+)$/)\n    if (ulMatch) {\n      blocks.push({\n        type: \"list\",\n        listType: \"unordered\",\n        text: ulMatch[2].trim(),\n        pageNumber,\n      })\n      i++\n      continue\n    }\n\n    // 5. 순서 리스트: 1.\n    const olMatch = line.match(/^(\\s*)\\d+\\.\\s+(.+)$/)\n    if (olMatch) {\n      blocks.push({\n        type: \"list\",\n        listType: \"ordered\",\n        text: olMatch[2].trim(),\n        pageNumber,\n      })\n      i++\n      continue\n    }\n\n    // 6. 일반 텍스트 — 구조적 행이 나올 때까지 병합\n    const paraLines: string[] = []\n    while (i < lines.length && lines[i].trim() !== \"\" && !isStructuralLine(lines[i])) {\n      paraLines.push(lines[i].trim())\n      i++\n    }\n    if (paraLines.length > 0) {\n      blocks.push({\n        type: \"paragraph\",\n        text: paraLines.join(\"\\n\"),\n        pageNumber,\n      })\n    }\n  }\n\n  return blocks\n}\n\n/**\n * 구조적 행 판별 — paragraph 병합 중단 트리거.\n */\nfunction isStructuralLine(line: string): boolean {\n  if (/^#{1,6}\\s+/.test(line)) return true\n  if (line.trim().startsWith(\"|\")) return true\n  if (/^[-*_]{3,}\\s*$/.test(line.trim())) return true\n  if (/^\\s*[-*+]\\s+/.test(line)) return true\n  if (/^\\s*\\d+\\.\\s+/.test(line)) return true\n  return false\n}\n\n/**\n * Markdown 테이블 행 배열을 IRTable로 변환.\n *\n * 구분선 행(|---|---|)은 제거 후 데이터 행만 파싱.\n * hasHeader: 구분선이 있었으면 true.\n */\nfunction parseMarkdownTable(lines: string[]): IRTable | null {\n  const hasSeparator = lines.some(line => /^\\|[\\s:|-]+\\|$/.test(line.trim()))\n\n  const rows: IRCell[][] = []\n  let maxCols = 0\n\n  for (const line of lines) {\n    // 구분선 행 스킵: |---|---| 패턴\n    if (/^\\|\\s*:?-+:?\\s*(\\|\\s*:?-+:?\\s*)+\\|?\\s*$/.test(line.trim())) continue\n\n    const parts = line.split(\"|\")\n    // 앞뒤 빈 요소 제거 (| 로 시작/종료하는 행)\n    const cells: IRCell[] = parts\n      .slice(1, parts[parts.length - 1].trim() === \"\" ? -1 : undefined)\n      .map(cell => ({\n        text: cell.trim(),\n        colSpan: 1,\n        rowSpan: 1,\n      }))\n\n    if (cells.length > 0) {\n      rows.push(cells)\n      maxCols = Math.max(maxCols, cells.length)\n    }\n  }\n\n  if (rows.length === 0) return null\n\n  // 열 수 통일 (부족한 셀은 빈 셀로 채움)\n  for (const row of rows) {\n    while (row.length < maxCols) {\n      row.push({ text: \"\", colSpan: 1, rowSpan: 1 })\n    }\n  }\n\n  return {\n    rows: rows.length,\n    cols: maxCols,\n    cells: rows,\n    hasHeader: hasSeparator && rows.length > 1,\n  }\n}\n"],"mappings":";;;AAoBO,SAAS,iBAAiB,UAAkB,YAA+B;AAChF,QAAM,SAAoB,CAAC;AAC3B,QAAM,QAAQ,SAAS,MAAM,IAAI;AACjC,MAAI,IAAI;AAER,SAAO,IAAI,MAAM,QAAQ;AACvB,UAAM,OAAO,MAAM,CAAC;AAGpB,QAAI,KAAK,KAAK,MAAM,IAAI;AACtB;AACA;AAAA,IACF;AAGA,UAAM,eAAe,KAAK,MAAM,mBAAmB;AACnD,QAAI,cAAc;AAChB,aAAO,KAAK;AAAA,QACV,MAAM;AAAA,QACN,OAAO,aAAa,CAAC,EAAE;AAAA,QACvB,MAAM,aAAa,CAAC,EAAE,KAAK;AAAA,QAC3B;AAAA,MACF,CAAC;AACD;AACA;AAAA,IACF;AAGA,QAAI,iBAAiB,KAAK,KAAK,KAAK,CAAC,GAAG;AACtC,aAAO,KAAK,EAAE,MAAM,aAAa,WAAW,CAAC;AAC7C;AACA;AAAA,IACF;AAGA,QAAI,KAAK,KAAK,EAAE,WAAW,GAAG,GAAG;AAC/B,YAAM,aAAuB,CAAC;AAC9B,aAAO,IAAI,MAAM,UAAU,MAAM,CAAC,EAAE,KAAK,EAAE,WAAW,GAAG,GAAG;AAC1D,mBAAW,KAAK,MAAM,CAAC,CAAC;AACxB;AAAA,MACF;AACA,YAAM,QAAQ,mBAAmB,UAAU;AAC3C,UAAI,OAAO;AACT,eAAO,KAAK,EAAE,MAAM,SAAS,OAAO,WAAW,CAAC;AAAA,MAClD;AACA;AAAA,IACF;AAGA,UAAM,UAAU,KAAK,MAAM,qBAAqB;AAChD,QAAI,SAAS;AACX,aAAO,KAAK;AAAA,QACV,MAAM;AAAA,QACN,UAAU;AAAA,QACV,MAAM,QAAQ,CAAC,EAAE,KAAK;AAAA,QACtB;AAAA,MACF,CAAC;AACD;AACA;AAAA,IACF;AAGA,UAAM,UAAU,KAAK,MAAM,qBAAqB;AAChD,QAAI,SAAS;AACX,aAAO,KAAK;AAAA,QACV,MAAM;AAAA,QACN,UAAU;AAAA,QACV,MAAM,QAAQ,CAAC,EAAE,KAAK;AAAA,QACtB;AAAA,MACF,CAAC;AACD;AACA;AAAA,IACF;AAGA,UAAM,YAAsB,CAAC;AAC7B,WAAO,IAAI,MAAM,UAAU,MAAM,CAAC,EAAE,KAAK,MAAM,MAAM,CAAC,iBAAiB,MAAM,CAAC,CAAC,GAAG;AAChF,gBAAU,KAAK,MAAM,CAAC,EAAE,KAAK,CAAC;AAC9B;AAAA,IACF;AACA,QAAI,UAAU,SAAS,GAAG;AACxB,aAAO,KAAK;AAAA,QACV,MAAM;AAAA,QACN,MAAM,UAAU,KAAK,IAAI;AAAA,QACzB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO;AACT;AAKA,SAAS,iBAAiB,MAAuB;AAC/C,MAAI,aAAa,KAAK,IAAI,EAAG,QAAO;AACpC,MAAI,KAAK,KAAK,EAAE,WAAW,GAAG,EAAG,QAAO;AACxC,MAAI,iBAAiB,KAAK,KAAK,KAAK,CAAC,EAAG,QAAO;AAC/C,MAAI,eAAe,KAAK,IAAI,EAAG,QAAO;AACtC,MAAI,eAAe,KAAK,IAAI,EAAG,QAAO;AACtC,SAAO;AACT;AAQA,SAAS,mBAAmB,OAAiC;AAC3D,QAAM,eAAe,MAAM,KAAK,UAAQ,iBAAiB,KAAK,KAAK,KAAK,CAAC,CAAC;AAE1E,QAAM,OAAmB,CAAC;AAC1B,MAAI,UAAU;AAEd,aAAW,QAAQ,OAAO;AAExB,QAAI,0CAA0C,KAAK,KAAK,KAAK,CAAC,EAAG;AAEjE,UAAM,QAAQ,KAAK,MAAM,GAAG;AAE5B,UAAM,QAAkB,MACrB,MAAM,GAAG,MAAM,MAAM,SAAS,CAAC,EAAE,KAAK,MAAM,KAAK,KAAK,MAAS,EAC/D,IAAI,WAAS;AAAA,MACZ,MAAM,KAAK,KAAK;AAAA,MAChB,SAAS;AAAA,MACT,SAAS;AAAA,IACX,EAAE;AAEJ,QAAI,MAAM,SAAS,GAAG;AACpB,WAAK,KAAK,KAAK;AACf,gBAAU,KAAK,IAAI,SAAS,MAAM,MAAM;AAAA,IAC1C;AAAA,EACF;AAEA,MAAI,KAAK,WAAW,EAAG,QAAO;AAG9B,aAAW,OAAO,MAAM;AACtB,WAAO,IAAI,SAAS,SAAS;AAC3B,UAAI,KAAK,EAAE,MAAM,IAAI,SAAS,GAAG,SAAS,EAAE,CAAC;AAAA,IAC/C;AAAA,EACF;AAEA,SAAO;AAAA,IACL,MAAM,KAAK;AAAA,IACX,MAAM;AAAA,IACN,OAAO;AAAA,IACP,WAAW,gBAAgB,KAAK,SAAS;AAAA,EAC3C;AACF;","names":[]}

package/dist/{tesseract-provider-WCVJWBUT.js → chunk-7FMKAV4P.js}

@@ -1,5 +1,4 @@
 #!/usr/bin/env node
-import "./chunk-ZWE3DS7E.js";
 
 // src/ocr/tesseract-provider.ts
 import { createWorker } from "tesseract.js";
@@ -49,8 +48,9 @@ async function createTesseractPoolProvider(concurrency) {
   };
   return provider;
 }
+
 export {
-  createTesseractPoolProvider,
-  createTesseractProvider
+  createTesseractProvider,
+  createTesseractPoolProvider
 };
-//# sourceMappingURL=tesseract-provider-WCVJWBUT.js.map
+//# sourceMappingURL=chunk-7FMKAV4P.js.map

package/dist/{tesseract-provider-WCVJWBUT.js.map → chunk-7FMKAV4P.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/ocr/tesseract-provider.ts"],"sourcesContent":["/**\n * Tesseract.js 기반 OCR 프로바이더\n *\n * tesseract.js는 bundled 의존성으로 별도 설치 불필요.\n * Vision LLM CLI가 없을 때의 최후 fallback으로 자동 사용.\n *\n * 특성:\n * - 순수 텍스트만 반환 (테이블/헤딩 구조 복원 불가)\n * - 한글 인식률 약 85-90% (깨끗한 이미지 기준)\n * - 완전 오프라인 동작 (API 키 불필요)\n */\n\nimport { createWorker } from \"tesseract.js\"\nimport type { OcrProvider } from \"../types.js\"\n\n/**\n * Tesseract.js OcrProvider 생성.\n *\n * 워커를 1회 생성하여 재사용 (매 페이지마다 초기화 방지).\n * 문서 처리 완료 후 terminate()로 워커 정리.\n *\n * @returns OcrProvider 함수 (+ terminate 메서드)\n */\nexport async function createTesseractProvider(): Promise<OcrProvider & { terminate: () => Promise<void> }> {\n  // kor+eng: 한글 + 영문 동시 인식 (한국 공문서 특성)\n  const worker = await createWorker(\"kor+eng\")\n  let terminated = false\n\n  const provider = async (\n    pageImage: Uint8Array,\n    _pageNumber: number,\n    _mimeType: \"image/png\"\n  ): Promise<string> => {\n    const { data } = await worker.recognize(pageImage)\n    return data.text\n  }\n\n  ;(provider as OcrProvider & { terminate: () => Promise<void> }).terminate = async () => {\n    if (!terminated) {\n      await worker.terminate()\n      terminated = true\n    }\n  }\n\n  return provider as OcrProvider & { terminate: () => Promise<void> }\n}\n\n/**\n * Tesseract.js 워커 풀 기반 병렬 OcrProvider 생성.\n *\n * N개 워커를 동시 생성하여 여러 페이지를 병렬로 OCR 처리.\n * acquire/release 패턴으로 동시 실행 안전성 보장.\n * 대량 이미지 기반 PDF(예: 300페이지) 처리 속도 향상에 효과적.\n *\n * @param concurrency - 동시 워커 수 (권장: CPU 코어 수)\n * @returns OcrProvider 함수 (+ terminate 메서드)\n */\nexport async function createTesseractPoolProvider(\n  concurrency: number\n): Promise<OcrProvider & { terminate: () => Promise<void> }> {\n  // N개 워커를 동시 초기화 (Promise.all로 병렬 생성)\n  const workers = await Promise.all(\n    Array.from({ length: concurrency }, () => createWorker(\"kor+eng\"))\n  )\n\n  // 유휴 워커 큐 (초기에는 모든 워커가 유휴 상태)\n  const idle: typeof workers = [...workers]\n  // 워커를 기다리는 콜백 큐\n  const waitQueue: ((w: (typeof workers)[0]) => void)[] = []\n\n  /** 유휴 워커를 획득 (없으면 대기) */\n  function acquire(): Promise<(typeof workers)[0]> {\n    if (idle.length > 0) return Promise.resolve(idle.pop()!)\n    return new Promise(resolve => waitQueue.push(resolve))\n  }\n\n  /** 사용 완료된 워커를 반환 (대기 중인 호출자에게 우선 전달) */\n  function release(w: (typeof workers)[0]): void {\n    if (waitQueue.length > 0) {\n      // 대기 중인 요청자에게 즉시 전달\n      waitQueue.shift()!(w)\n    } else {\n      idle.push(w)\n    }\n  }\n\n  // 병렬 호출 안전 프로바이더 — 각 호출이 acquire/release로 워커를 독점 사용\n  const provider = async (\n    pageImage: Uint8Array,\n    _pageNumber: number,\n    _mimeType: \"image/png\"\n  ): Promise<string> => {\n    const w = await acquire()\n    try {\n      const { data } = await w.recognize(pageImage)\n      return data.text\n    } finally {\n      // 성공/실패 무관하게 워커 반환\n      release(w)\n    }\n  }\n\n  // 모든 워커 정리 메서드\n  ;(provider as OcrProvider & { terminate: () => Promise<void> }).terminate = async () => {\n    await Promise.all(workers.map(w => w.terminate()))\n  }\n\n  return provider as OcrProvider & { terminate: () => Promise<void> }\n}\n"],"mappings":";;;;AAYA,SAAS,oBAAoB;AAW7B,eAAsB,0BAAqF;AAEzG,QAAM,SAAS,MAAM,aAAa,SAAS;AAC3C,MAAI,aAAa;AAEjB,QAAM,WAAW,OACf,WACA,aACA,cACoB;AACpB,UAAM,EAAE,KAAK,IAAI,MAAM,OAAO,UAAU,SAAS;AACjD,WAAO,KAAK;AAAA,EACd;AAEC,EAAC,SAA8D,YAAY,YAAY;AACtF,QAAI,CAAC,YAAY;AACf,YAAM,OAAO,UAAU;AACvB,mBAAa;AAAA,IACf;AAAA,EACF;AAEA,SAAO;AACT;AAYA,eAAsB,4BACpB,aAC2D;AAE3D,QAAM,UAAU,MAAM,QAAQ;AAAA,IAC5B,MAAM,KAAK,EAAE,QAAQ,YAAY,GAAG,MAAM,aAAa,SAAS,CAAC;AAAA,EACnE;AAGA,QAAM,OAAuB,CAAC,GAAG,OAAO;AAExC,QAAM,YAAkD,CAAC;AAGzD,WAAS,UAAwC;AAC/C,QAAI,KAAK,SAAS,EAAG,QAAO,QAAQ,QAAQ,KAAK,IAAI,CAAE;AACvD,WAAO,IAAI,QAAQ,aAAW,UAAU,KAAK,OAAO,CAAC;AAAA,EACvD;AAGA,WAAS,QAAQ,GAA8B;AAC7C,QAAI,UAAU,SAAS,GAAG;AAExB,gBAAU,MAAM,EAAG,CAAC;AAAA,IACtB,OAAO;AACL,WAAK,KAAK,CAAC;AAAA,IACb;AAAA,EACF;AAGA,QAAM,WAAW,OACf,WACA,aACA,cACoB;AACpB,UAAM,IAAI,MAAM,QAAQ;AACxB,QAAI;AACF,YAAM,EAAE,KAAK,IAAI,MAAM,EAAE,UAAU,SAAS;AAC5C,aAAO,KAAK;AAAA,IACd,UAAE;AAEA,cAAQ,CAAC;AAAA,IACX;AAAA,EACF;AAGC,EAAC,SAA8D,YAAY,YAAY;AACtF,UAAM,QAAQ,IAAI,QAAQ,IAAI,OAAK,EAAE,UAAU,CAAC,CAAC;AAAA,EACnD;AAEA,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/ocr/tesseract-provider.ts"],"sourcesContent":["/**\n * Tesseract.js 기반 OCR 프로바이더\n *\n * tesseract.js는 bundled 의존성으로 별도 설치 불필요.\n * Vision LLM CLI가 없을 때의 최후 fallback으로 자동 사용.\n *\n * 특성:\n * - 순수 텍스트만 반환 (테이블/헤딩 구조 복원 불가)\n * - 한글 인식률 약 85-90% (깨끗한 이미지 기준)\n * - 완전 오프라인 동작 (API 키 불필요)\n */\n\nimport { createWorker } from \"tesseract.js\"\nimport type { OcrProvider } from \"../types.js\"\n\n/**\n * Tesseract.js OcrProvider 생성.\n *\n * 워커를 1회 생성하여 재사용 (매 페이지마다 초기화 방지).\n * 문서 처리 완료 후 terminate()로 워커 정리.\n *\n * @returns OcrProvider 함수 (+ terminate 메서드)\n */\nexport async function createTesseractProvider(): Promise<OcrProvider & { terminate: () => Promise<void> }> {\n  // kor+eng: 한글 + 영문 동시 인식 (한국 공문서 특성)\n  const worker = await createWorker(\"kor+eng\")\n  let terminated = false\n\n  const provider = async (\n    pageImage: Uint8Array,\n    _pageNumber: number,\n    _mimeType: \"image/png\"\n  ): Promise<string> => {\n    const { data } = await worker.recognize(pageImage)\n    return data.text\n  }\n\n  ;(provider as OcrProvider & { terminate: () => Promise<void> }).terminate = async () => {\n    if (!terminated) {\n      await worker.terminate()\n      terminated = true\n    }\n  }\n\n  return provider as OcrProvider & { terminate: () => Promise<void> }\n}\n\n/**\n * Tesseract.js 워커 풀 기반 병렬 OcrProvider 생성.\n *\n * N개 워커를 동시 생성하여 여러 페이지를 병렬로 OCR 처리.\n * acquire/release 패턴으로 동시 실행 안전성 보장.\n * 대량 이미지 기반 PDF(예: 300페이지) 처리 속도 향상에 효과적.\n *\n * @param concurrency - 동시 워커 수 (권장: CPU 코어 수)\n * @returns OcrProvider 함수 (+ terminate 메서드)\n */\nexport async function createTesseractPoolProvider(\n  concurrency: number\n): Promise<OcrProvider & { terminate: () => Promise<void> }> {\n  // N개 워커를 동시 초기화 (Promise.all로 병렬 생성)\n  const workers = await Promise.all(\n    Array.from({ length: concurrency }, () => createWorker(\"kor+eng\"))\n  )\n\n  // 유휴 워커 큐 (초기에는 모든 워커가 유휴 상태)\n  const idle: typeof workers = [...workers]\n  // 워커를 기다리는 콜백 큐\n  const waitQueue: ((w: (typeof workers)[0]) => void)[] = []\n\n  /** 유휴 워커를 획득 (없으면 대기) */\n  function acquire(): Promise<(typeof workers)[0]> {\n    if (idle.length > 0) return Promise.resolve(idle.pop()!)\n    return new Promise(resolve => waitQueue.push(resolve))\n  }\n\n  /** 사용 완료된 워커를 반환 (대기 중인 호출자에게 우선 전달) */\n  function release(w: (typeof workers)[0]): void {\n    if (waitQueue.length > 0) {\n      // 대기 중인 요청자에게 즉시 전달\n      waitQueue.shift()!(w)\n    } else {\n      idle.push(w)\n    }\n  }\n\n  // 병렬 호출 안전 프로바이더 — 각 호출이 acquire/release로 워커를 독점 사용\n  const provider = async (\n    pageImage: Uint8Array,\n    _pageNumber: number,\n    _mimeType: \"image/png\"\n  ): Promise<string> => {\n    const w = await acquire()\n    try {\n      const { data } = await w.recognize(pageImage)\n      return data.text\n    } finally {\n      // 성공/실패 무관하게 워커 반환\n      release(w)\n    }\n  }\n\n  // 모든 워커 정리 메서드\n  ;(provider as OcrProvider & { terminate: () => Promise<void> }).terminate = async () => {\n    await Promise.all(workers.map(w => w.terminate()))\n  }\n\n  return provider as OcrProvider & { terminate: () => Promise<void> }\n}\n"],"mappings":";;;AAYA,SAAS,oBAAoB;AAW7B,eAAsB,0BAAqF;AAEzG,QAAM,SAAS,MAAM,aAAa,SAAS;AAC3C,MAAI,aAAa;AAEjB,QAAM,WAAW,OACf,WACA,aACA,cACoB;AACpB,UAAM,EAAE,KAAK,IAAI,MAAM,OAAO,UAAU,SAAS;AACjD,WAAO,KAAK;AAAA,EACd;AAEC,EAAC,SAA8D,YAAY,YAAY;AACtF,QAAI,CAAC,YAAY;AACf,YAAM,OAAO,UAAU;AACvB,mBAAa;AAAA,IACf;AAAA,EACF;AAEA,SAAO;AACT;AAYA,eAAsB,4BACpB,aAC2D;AAE3D,QAAM,UAAU,MAAM,QAAQ;AAAA,IAC5B,MAAM,KAAK,EAAE,QAAQ,YAAY,GAAG,MAAM,aAAa,SAAS,CAAC;AAAA,EACnE;AAGA,QAAM,OAAuB,CAAC,GAAG,OAAO;AAExC,QAAM,YAAkD,CAAC;AAGzD,WAAS,UAAwC;AAC/C,QAAI,KAAK,SAAS,EAAG,QAAO,QAAQ,QAAQ,KAAK,IAAI,CAAE;AACvD,WAAO,IAAI,QAAQ,aAAW,UAAU,KAAK,OAAO,CAAC;AAAA,EACvD;AAGA,WAAS,QAAQ,GAA8B;AAC7C,QAAI,UAAU,SAAS,GAAG;AAExB,gBAAU,MAAM,EAAG,CAAC;AAAA,IACtB,OAAO;AACL,WAAK,KAAK,CAAC;AAAA,IACb;AAAA,EACF;AAGA,QAAM,WAAW,OACf,WACA,aACA,cACoB;AACpB,UAAM,IAAI,MAAM,QAAQ;AACxB,QAAI;AACF,YAAM,EAAE,KAAK,IAAI,MAAM,EAAE,UAAU,SAAS;AAC5C,aAAO,KAAK;AAAA,IACd,UAAE;AAEA,cAAQ,CAAC;AAAA,IACX;AAAA,EACF;AAGC,EAAC,SAA8D,YAAY,YAAY;AACtF,UAAM,QAAQ,IAAI,QAAQ,IAAI,OAAK,EAAE,UAAU,CAAC,CAAC;AAAA,EACnD;AAEA,SAAO;AACT;","names":[]}

package/dist/chunk-JOGAFNIL.js

@@ -0,0 +1,153 @@
+#!/usr/bin/env node
+
+// src/ocr/cli-provider.ts
+import { spawnSync } from "child_process";
+import { writeFileSync, readFileSync, unlinkSync, mkdirSync } from "fs";
+import { join } from "path";
+import { tmpdir } from "os";
+var 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`;
+var _tempDir = null;
+function getTempDir() {
+  if (!_tempDir) {
+    _tempDir = join(process.cwd(), "_kordoc_ocr_tmp");
+    mkdirSync(_tempDir, { recursive: true });
+  }
+  return _tempDir;
+}
+function createCliOcrProvider(mode) {
+  return async (pageImage, pageNumber) => {
+    const tempPath = join(getTempDir(), `page-${pageNumber}.png`);
+    try {
+      writeFileSync(tempPath, pageImage);
+      let output;
+      if (mode === "ollama") {
+        output = await callOllamaApi(tempPath);
+      } else {
+        output = callCli(mode, tempPath);
+      }
+      return { markdown: stripCodeFence(output.trim()) };
+    } finally {
+      try {
+        unlinkSync(tempPath);
+      } catch {
+      }
+    }
+  };
+}
+function callCli(mode, imagePath) {
+  if (mode === "codex") {
+    return callCodexCli(imagePath);
+  }
+  const args = buildCliArgs(mode, imagePath);
+  const result = spawnSync(mode, args, {
+    encoding: "utf-8",
+    timeout: 6e5,
+    maxBuffer: 10 * 1024 * 1024,
+    // claude: /tmp에서 실행하여 프로젝트 CLAUDE.md의 규칙 간섭 방지
+    ...mode === "claude" ? { cwd: tmpdir() } : {}
+  });
+  if (result.error) {
+    throw new Error(`${mode} CLI \uC2E4\uD589 \uC2E4\uD328: ${result.error.message}`);
+  }
+  if (result.status !== 0) {
+    const errMsg = result.stderr?.trim() || `exit code ${result.status}`;
+    throw new Error(`${mode} OCR \uC2E4\uD328: ${errMsg}`);
+  }
+  return result.stdout || "";
+}
+function callCodexCli(imagePath) {
+  const outPath = join(tmpdir(), `kordoc-codex-out-${Date.now()}.txt`);
+  try {
+    const args = ["exec", OCR_PROMPT, "--image", imagePath, "--output-last-message", outPath];
+    const model = process.env.KORDOC_CODEX_MODEL;
+    if (model) args.push("--model", model);
+    const result = spawnSync("codex", args, {
+      encoding: "utf-8",
+      timeout: 18e4,
+      maxBuffer: 10 * 1024 * 1024,
+      input: ""
+      // stdin EOF 즉시 전달 (대화형 입력 차단)
+    });
+    if (result.error) {
+      throw new Error(`codex CLI \uC2E4\uD589 \uC2E4\uD328: ${result.error.message}`);
+    }
+    if (result.status !== 0) {
+      const errMsg = result.stderr?.trim() || `exit code ${result.status}`;
+      throw new Error(`codex OCR \uC2E4\uD328: ${errMsg}`);
+    }
+    try {
+      return readFileSync(outPath, "utf-8");
+    } catch {
+      return result.stdout || "";
+    }
+  } finally {
+    try {
+      unlinkSync(outPath);
+    } catch {
+    }
+  }
+}
+function buildCliArgs(mode, imagePath) {
+  const promptWithImage = `${OCR_PROMPT}
+
+\uC774\uBBF8\uC9C0: @${imagePath}`;
+  switch (mode) {
+    case "gemini": {
+      const args = ["--prompt", promptWithImage, "--yolo"];
+      const model = process.env.KORDOC_GEMINI_MODEL;
+      if (model) args.push("--model", model);
+      return args;
+    }
+    case "claude": {
+      const args = ["--print", promptWithImage];
+      const model = process.env.KORDOC_CLAUDE_MODEL;
+      if (model) args.push("--model", model);
+      return args;
+    }
+    default:
+      throw new Error(`\uC9C0\uC6D0\uD558\uC9C0 \uC54A\uB294 CLI: ${mode}`);
+  }
+}
+async function callOllamaApi(imagePath) {
+  const { readFileSync: readFileSync2 } = await import("fs");
+  const imageBase64 = readFileSync2(imagePath).toString("base64");
+  const model = process.env.KORDOC_OLLAMA_MODEL || "qwen3-vl:8b";
+  const host = process.env.KORDOC_OLLAMA_HOST || "http://localhost:11434";
+  const timeoutMs = Number(process.env.KORDOC_OLLAMA_TIMEOUT) || 12e4;
+  const response = await fetch(`${host}/api/chat`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      model,
+      messages: [{
+        role: "user",
+        content: OCR_PROMPT,
+        images: [imageBase64]
+      }],
+      stream: false
+    }),
+    signal: AbortSignal.timeout(timeoutMs)
+  });
+  if (!response.ok) {
+    throw new Error(`Ollama API \uC624\uB958: ${response.status} ${response.statusText}`);
+  }
+  const data = await response.json();
+  return data.message?.content || "";
+}
+function stripCodeFence(text) {
+  const match = text.match(/^```(?:markdown|md)?\s*([\s\S]*?)```\s*$/m);
+  return match ? match[1].trim() : text;
+}
+
+export {
+  createCliOcrProvider
+};
+//# sourceMappingURL=chunk-JOGAFNIL.js.map

package/dist/chunk-JOGAFNIL.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/ocr/cli-provider.ts"],"sourcesContent":["/**\n * CLI 기반 OCR 프로바이더\n *\n * gemini / claude / codex / ollama CLI를 subprocess로 호출하여\n * PDF 페이지 이미지를 Markdown으로 변환.\n *\n * 이미지 전달 방식:\n * - gemini:  -p \"프롬프트 @이미지경로\"  (@ 파일 참조)\n * - claude:  -p \"프롬프트 @이미지경로\"  (@ 파일 참조, --print 모드)\n * - codex:   exec -i 이미지경로 \"프롬프트\"  (-i/--image 플래그)\n * - ollama:  REST API (localhost:11434) — CLI는 이미지 입력 미지원\n */\n\nimport { spawnSync } from \"child_process\"\nimport { writeFileSync, readFileSync, unlinkSync, mkdirSync } from \"fs\"\nimport { join } from \"path\"\nimport { tmpdir } from \"os\"\nimport type { OcrMode, StructuredOcrResult } from \"../types.js\"\n\n/** OCR 프롬프트 — 모든 CLI 공통 */\nconst OCR_PROMPT = `이 PDF 페이지 이미지에서 텍스트와 테이블을 추출하여 순수 Markdown으로 변환하세요.\n규칙:\n- 테이블은 Markdown 테이블 문법 사용 (| 구분, |---|---| 헤더 구분선 포함)\n- 병합된 셀은 해당 위치에 내용 기재\n- 헤딩은 글자 크기에 따라 ## ~ ###### 사용\n- 리스트는 - 또는 1. 사용\n- 이미지, 도형 등 비텍스트 요소는 무시\n- 원문의 읽기 순서와 구조를 유지\n- \\`\\`\\`로 감싸지 말고 순수 Markdown만 출력`\n\n/** 임시 디렉토리 (프로세스당 1회 생성)\n *\n * gemini CLI는 /tmp/ 등 시스템 임시 디렉토리를 워크스페이스 외부로 간주하여\n * @파일참조 시 접근을 거부할 수 있음. cwd 하위 폴더를 사용하면 모든 CLI에서 접근 가능.\n *\n * ⚠️ .gitignore에 포함된 경로(예: .kordoc-tmp/)는 gemini CLI가 무시하므로\n * 반드시 gitignore되지 않는 이름 사용. 파일은 try/finally로 즉시 정리.\n */\nlet _tempDir: string | null = null\nfunction getTempDir(): string {\n  if (!_tempDir) {\n    _tempDir = join(process.cwd(), \"_kordoc_ocr_tmp\")\n    mkdirSync(_tempDir, { recursive: true })\n  }\n  return _tempDir\n}\n\n/**\n * CLI OcrProvider 생성.\n *\n * @param mode - 사용할 CLI (gemini, claude, codex, ollama)\n * @returns OcrProvider 함수 (StructuredOcrResult 반환)\n */\nexport function createCliOcrProvider(\n  mode: Exclude<OcrMode, \"auto\" | \"off\" | \"tesseract\">\n): (pageImage: Uint8Array, pageNumber: number, mimeType: \"image/png\") => Promise<StructuredOcrResult> {\n  return async (pageImage: Uint8Array, pageNumber: number): Promise<StructuredOcrResult> => {\n    const tempPath = join(getTempDir(), `page-${pageNumber}.png`)\n\n    try {\n      writeFileSync(tempPath, pageImage)\n\n      let output: string\n      if (mode === \"ollama\") {\n        output = await callOllamaApi(tempPath)\n      } else {\n        output = callCli(mode, tempPath)\n      }\n\n      return { markdown: stripCodeFence(output.trim()) }\n    } finally {\n      try { unlinkSync(tempPath) } catch { /* 임시 파일 정리 실패 무시 */ }\n    }\n  }\n}\n\n/**\n * CLI 실행 — gemini / claude / codex\n *\n * @throws CLI 실행 실패 또는 타임아웃(180초) 시 Error\n */\nfunction callCli(mode: string, imagePath: string): string {\n  // codex는 --output-last-message로 대화 헤더 없는 깔끔한 출력 사용\n  if (mode === \"codex\") {\n    return callCodexCli(imagePath)\n  }\n\n  const args = buildCliArgs(mode, imagePath)\n\n  const result = spawnSync(mode, args, {\n    encoding: \"utf-8\",\n    timeout: 600_000,\n    maxBuffer: 10 * 1024 * 1024,\n    // claude: /tmp에서 실행하여 프로젝트 CLAUDE.md의 규칙 간섭 방지\n    ...(mode === \"claude\" ? { cwd: tmpdir() } : {}),\n  })\n\n  if (result.error) {\n    throw new Error(`${mode} CLI 실행 실패: ${result.error.message}`)\n  }\n  if (result.status !== 0) {\n    const errMsg = result.stderr?.trim() || `exit code ${result.status}`\n    throw new Error(`${mode} OCR 실패: ${errMsg}`)\n  }\n\n  return result.stdout || \"\"\n}\n\n/**\n * codex exec 실행 — --output-last-message로 대화 헤더 없는 깔끔한 출력.\n * 인자 순서: `codex exec <prompt> --image <file> --output-last-message <outfile>`\n */\nfunction callCodexCli(imagePath: string): string {\n  // 출력 파일은 /tmp/ 사용 — codex sandbox는 cwd 내 쓰기를 막을 수 있음\n  const outPath = join(tmpdir(), `kordoc-codex-out-${Date.now()}.txt`)\n  try {\n    const args = [\"exec\", OCR_PROMPT, \"--image\", imagePath, \"--output-last-message\", outPath]\n    const model = process.env.KORDOC_CODEX_MODEL\n    if (model) args.push(\"--model\", model)\n\n    const result = spawnSync(\"codex\", args, {\n      encoding: \"utf-8\",\n      timeout: 180_000,\n      maxBuffer: 10 * 1024 * 1024,\n      input: \"\",  // stdin EOF 즉시 전달 (대화형 입력 차단)\n    })\n\n    if (result.error) {\n      throw new Error(`codex CLI 실행 실패: ${result.error.message}`)\n    }\n    if (result.status !== 0) {\n      const errMsg = result.stderr?.trim() || `exit code ${result.status}`\n      throw new Error(`codex OCR 실패: ${errMsg}`)\n    }\n\n    // --output-last-message 파일에서 읽기 (없으면 stdout 폴백)\n    try {\n      return readFileSync(outPath, \"utf-8\")\n    } catch {\n      return result.stdout || \"\"\n    }\n  } finally {\n    try { unlinkSync(outPath) } catch { /* 무시 */ }\n  }\n}\n\n/**\n * CLI별 인자 배열 생성.\n *\n * gemini:  [\"--prompt\", \"프롬프트 @이미지경로\", \"--yolo\"]\n *   - -y/--yolo: 자동 승인 (OCR은 도구 사용 없으므로 실질적 영향 없음)\n *   - @ 파일 참조로 이미지를 컨텍스트에 포함\n *\n * claude:  [\"--print\", \"프롬프트 @이미지경로\"]\n *   - --print(-p): 비대화형 출력 모드\n *   - @ 파일 참조로 이미지를 컨텍스트에 포함\n *\n * codex: callCodexCli()에서 별도 처리\n *   - `codex exec <prompt> --image <file> --output-last-message <outfile>`\n *   - 프롬프트가 --image보다 앞에 위치해야 함 (인자 순서 중요)\n *\n * ⚠️ CLI 버전에 따라 문법이 다를 수 있음. 업데이트 시 --help 재확인 필요.\n */\nfunction buildCliArgs(mode: string, imagePath: string): string[] {\n  const promptWithImage = `${OCR_PROMPT}\n\n이미지: @${imagePath}`\n\n  switch (mode) {\n    case \"gemini\": {\n      const args = [\"--prompt\", promptWithImage, \"--yolo\"]\n      const model = process.env.KORDOC_GEMINI_MODEL\n      if (model) args.push(\"--model\", model)\n      return args\n    }\n\n    case \"claude\": {\n      const args = [\"--print\", promptWithImage]\n      const model = process.env.KORDOC_CLAUDE_MODEL\n      if (model) args.push(\"--model\", model)\n      return args\n    }\n\n    default:\n      throw new Error(`지원하지 않는 CLI: ${mode}`)\n  }\n}\n\n/**\n * Ollama REST API 호출 — CLI는 이미지 입력을 지원하지 않으므로 API 직접 사용.\n *\n * 기본 모델: KORDOC_OLLAMA_MODEL 환경변수 또는 \"gemma4:27b\"\n * 기본 호스트: KORDOC_OLLAMA_HOST 환경변수 또는 \"http://localhost:11434\"\n *\n * @throws Ollama 서버 미실행 또는 응답 오류 시 Error\n */\nasync function callOllamaApi(imagePath: string): Promise<string> {\n  const { readFileSync } = await import(\"fs\")\n  const imageBase64 = readFileSync(imagePath).toString(\"base64\")\n\n  const model = process.env.KORDOC_OLLAMA_MODEL || \"qwen3-vl:8b\"\n  const host = process.env.KORDOC_OLLAMA_HOST || \"http://localhost:11434\"\n  const timeoutMs = Number(process.env.KORDOC_OLLAMA_TIMEOUT) || 120_000\n\n  const response = await fetch(`${host}/api/chat`, {\n    method: \"POST\",\n    headers: { \"Content-Type\": \"application/json\" },\n    body: JSON.stringify({\n      model,\n      messages: [{\n        role: \"user\",\n        content: OCR_PROMPT,\n        images: [imageBase64],\n      }],\n      stream: false,\n    }),\n    signal: AbortSignal.timeout(timeoutMs),\n  })\n\n  if (!response.ok) {\n    throw new Error(`Ollama API 오류: ${response.status} ${response.statusText}`)\n  }\n\n  const data = await response.json() as { message?: { content?: string } }\n  return data.message?.content || \"\"\n}\n\n/**\n * LLM 출력에서 코드 펜스 제거.\n * LLM이 가끔 결과를 ```markdown ... ``` 으로 감싸는 경우 처리.\n */\nfunction stripCodeFence(text: string): string {\n  const match = text.match(/^```(?:markdown|md)?\\s*([\\s\\S]*?)```\\s*$/m)\n  return match ? match[1].trim() : text\n}\n"],"mappings":";;;AAaA,SAAS,iBAAiB;AAC1B,SAAS,eAAe,cAAc,YAAY,iBAAiB;AACnE,SAAS,YAAY;AACrB,SAAS,cAAc;AAIvB,IAAM,aAAa;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAkBnB,IAAI,WAA0B;AAC9B,SAAS,aAAqB;AAC5B,MAAI,CAAC,UAAU;AACb,eAAW,KAAK,QAAQ,IAAI,GAAG,iBAAiB;AAChD,cAAU,UAAU,EAAE,WAAW,KAAK,CAAC;AAAA,EACzC;AACA,SAAO;AACT;AAQO,SAAS,qBACd,MACoG;AACpG,SAAO,OAAO,WAAuB,eAAqD;AACxF,UAAM,WAAW,KAAK,WAAW,GAAG,QAAQ,UAAU,MAAM;AAE5D,QAAI;AACF,oBAAc,UAAU,SAAS;AAEjC,UAAI;AACJ,UAAI,SAAS,UAAU;AACrB,iBAAS,MAAM,cAAc,QAAQ;AAAA,MACvC,OAAO;AACL,iBAAS,QAAQ,MAAM,QAAQ;AAAA,MACjC;AAEA,aAAO,EAAE,UAAU,eAAe,OAAO,KAAK,CAAC,EAAE;AAAA,IACnD,UAAE;AACA,UAAI;AAAE,mBAAW,QAAQ;AAAA,MAAE,QAAQ;AAAA,MAAuB;AAAA,IAC5D;AAAA,EACF;AACF;AAOA,SAAS,QAAQ,MAAc,WAA2B;AAExD,MAAI,SAAS,SAAS;AACpB,WAAO,aAAa,SAAS;AAAA,EAC/B;AAEA,QAAM,OAAO,aAAa,MAAM,SAAS;AAEzC,QAAM,SAAS,UAAU,MAAM,MAAM;AAAA,IACnC,UAAU;AAAA,IACV,SAAS;AAAA,IACT,WAAW,KAAK,OAAO;AAAA;AAAA,IAEvB,GAAI,SAAS,WAAW,EAAE,KAAK,OAAO,EAAE,IAAI,CAAC;AAAA,EAC/C,CAAC;AAED,MAAI,OAAO,OAAO;AAChB,UAAM,IAAI,MAAM,GAAG,IAAI,mCAAe,OAAO,MAAM,OAAO,EAAE;AAAA,EAC9D;AACA,MAAI,OAAO,WAAW,GAAG;AACvB,UAAM,SAAS,OAAO,QAAQ,KAAK,KAAK,aAAa,OAAO,MAAM;AAClE,UAAM,IAAI,MAAM,GAAG,IAAI,sBAAY,MAAM,EAAE;AAAA,EAC7C;AAEA,SAAO,OAAO,UAAU;AAC1B;AAMA,SAAS,aAAa,WAA2B;AAE/C,QAAM,UAAU,KAAK,OAAO,GAAG,oBAAoB,KAAK,IAAI,CAAC,MAAM;AACnE,MAAI;AACF,UAAM,OAAO,CAAC,QAAQ,YAAY,WAAW,WAAW,yBAAyB,OAAO;AACxF,UAAM,QAAQ,QAAQ,IAAI;AAC1B,QAAI,MAAO,MAAK,KAAK,WAAW,KAAK;AAErC,UAAM,SAAS,UAAU,SAAS,MAAM;AAAA,MACtC,UAAU;AAAA,MACV,SAAS;AAAA,MACT,WAAW,KAAK,OAAO;AAAA,MACvB,OAAO;AAAA;AAAA,IACT,CAAC;AAED,QAAI,OAAO,OAAO;AAChB,YAAM,IAAI,MAAM,wCAAoB,OAAO,MAAM,OAAO,EAAE;AAAA,IAC5D;AACA,QAAI,OAAO,WAAW,GAAG;AACvB,YAAM,SAAS,OAAO,QAAQ,KAAK,KAAK,aAAa,OAAO,MAAM;AAClE,YAAM,IAAI,MAAM,2BAAiB,MAAM,EAAE;AAAA,IAC3C;AAGA,QAAI;AACF,aAAO,aAAa,SAAS,OAAO;AAAA,IACtC,QAAQ;AACN,aAAO,OAAO,UAAU;AAAA,IAC1B;AAAA,EACF,UAAE;AACA,QAAI;AAAE,iBAAW,OAAO;AAAA,IAAE,QAAQ;AAAA,IAAW;AAAA,EAC/C;AACF;AAmBA,SAAS,aAAa,MAAc,WAA6B;AAC/D,QAAM,kBAAkB,GAAG,UAAU;AAAA;AAAA,uBAE/B,SAAS;AAEf,UAAQ,MAAM;AAAA,IACZ,KAAK,UAAU;AACb,YAAM,OAAO,CAAC,YAAY,iBAAiB,QAAQ;AACnD,YAAM,QAAQ,QAAQ,IAAI;AAC1B,UAAI,MAAO,MAAK,KAAK,WAAW,KAAK;AACrC,aAAO;AAAA,IACT;AAAA,IAEA,KAAK,UAAU;AACb,YAAM,OAAO,CAAC,WAAW,eAAe;AACxC,YAAM,QAAQ,QAAQ,IAAI;AAC1B,UAAI,MAAO,MAAK,KAAK,WAAW,KAAK;AACrC,aAAO;AAAA,IACT;AAAA,IAEA;AACE,YAAM,IAAI,MAAM,8CAAgB,IAAI,EAAE;AAAA,EAC1C;AACF;AAUA,eAAe,cAAc,WAAoC;AAC/D,QAAM,EAAE,cAAAA,cAAa,IAAI,MAAM,OAAO,IAAI;AAC1C,QAAM,cAAcA,cAAa,SAAS,EAAE,SAAS,QAAQ;AAE7D,QAAM,QAAQ,QAAQ,IAAI,uBAAuB;AACjD,QAAM,OAAO,QAAQ,IAAI,sBAAsB;AAC/C,QAAM,YAAY,OAAO,QAAQ,IAAI,qBAAqB,KAAK;AAE/D,QAAM,WAAW,MAAM,MAAM,GAAG,IAAI,aAAa;AAAA,IAC/C,QAAQ;AAAA,IACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,IAC9C,MAAM,KAAK,UAAU;AAAA,MACnB;AAAA,MACA,UAAU,CAAC;AAAA,QACT,MAAM;AAAA,QACN,SAAS;AAAA,QACT,QAAQ,CAAC,WAAW;AAAA,MACtB,CAAC;AAAA,MACD,QAAQ;AAAA,IACV,CAAC;AAAA,IACD,QAAQ,YAAY,QAAQ,SAAS;AAAA,EACvC,CAAC;AAED,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,IAAI,MAAM,4BAAkB,SAAS,MAAM,IAAI,SAAS,UAAU,EAAE;AAAA,EAC5E;AAEA,QAAM,OAAO,MAAM,SAAS,KAAK;AACjC,SAAO,KAAK,SAAS,WAAW;AAClC;AAMA,SAAS,eAAe,MAAsB;AAC5C,QAAM,QAAQ,KAAK,MAAM,2CAA2C;AACpE,SAAO,QAAQ,MAAM,CAAC,EAAE,KAAK,IAAI;AACnC;","names":["readFileSync"]}

package/dist/{chunk-W5KUC23B.js → chunk-NU3KFVVZ.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 // src/utils.ts
-var VERSION = true ? "2.3.1" : "0.0.0-dev";
+var VERSION = true ? "2.3.3" : "0.0.0-dev";
 function toArrayBuffer(buf) {
   if (buf.byteOffset === 0 && buf.byteLength === buf.buffer.byteLength) {
     return buf.buffer;
@@ -90,4 +90,4 @@ export {
   sanitizeHref,
   classifyError
 };
-//# sourceMappingURL=chunk-W5KUC23B.js.map
+//# sourceMappingURL=chunk-NU3KFVVZ.js.map