#!/usr/bin/env python3 """KoOCR-Bench v1.2 API runs with block boxes (owner 2026-10-05: every engine must be asked for coordinates). Output per block: <|det|>TYPE [x0, y0, x1, y1]<|/det|>CONTENT, coordinates normalised to 0..999 of the page. gemini-* : PROMPT below (published verbatim in the benchmark README); temperature 0, lowest thinking setting (3.1: thinkingBudget 0; 3.5 rejects budget 0, so thinkingLevel "minimal"). A det header the model closed with <|det|> or <|/det/> instead of <|/det|> is repaired (CLOSE_FIX) so the block text is not swallowed by the next tag. mistral : mistral-ocr-latest, table_format=html; uses the response's `blocks` (top_left/bottom_right px) in API order, table blocks replaced by the returned HTML table. clova : CLOVA OCR V2 general + table detection; words grouped into lines by lineBreak, line box = union of word boundingPoly; lines whose centre falls inside a detected table are replaced by that table (HTML) as one block; blocks ordered by top y then x. Raw API responses are kept next to OUT (OUT.raw.jsonl). Keys/endpoints come from env: GEMINI_API_KEY, MISTRAL_API_KEY, CLOVA_OCR_URL, CLOVA_OCR_SECRET. Usage: api_run_v3.py ENGINE MANIFEST_JSONL IMG_ROOT OUT_JSONL [THREADS]""" import base64, html, io, json, os, re, sys, time, uuid, urllib.request from concurrent.futures import ThreadPoolExecutor from pathlib import Path from PIL import Image eng, man, root, out = sys.argv[1], sys.argv[2], Path(sys.argv[3]), Path(sys.argv[4]) threads = int(sys.argv[5]) if len(sys.argv) > 5 else 8 rawf = Path(str(out) + ".raw.jsonl") M = [json.loads(l) for l in open(man)] done = ( {json.loads(l)["id"] for l in out.open() if not json.loads(l).get("error")} if out.exists() else set() ) todo = [r for r in M if r["id"] not in done] PROMPT = "\n".join( [ "이 문서 페이지를 OCR 하라. 페이지를 읽는 순서대로 블록(문단, 제목, 표, 그림, 머리말, 꼬리말, 쪽번호, 수식, 캡션) 단위로 나눠 출력하라.", "- 각 블록은 반드시 `<|det|>종류 [x0, y0, x1, y1]<|/det|>내용` 형식으로 한 블록씩 출력하라.", "- 종류는 title, text, table, image, header, footer, page_number, equation, caption 중 하나다.", "- 좌표는 페이지 왼쪽 위가 (0, 0), 오른쪽 아래가 (999, 999)인 정수다. x0,y0은 블록 왼쪽 위, x1,y1은 오른쪽 아래다.", "- 원문 표현은 그대로 보존하라(요약·바꿔쓰기·날조 금지). 보이는 텍스트만.", "- 한 문단 안에서 줄이 끊긴 것은 이어 붙여라.", "- 표는 HTML
로 출력하라. 셀 병합은 rowspan/colspan으로 표시하라. Markdown 표 금지.", "- 그림은 내용 없이 `<|det|>image [x0, y0, x1, y1]<|/det|>`만 출력하라.", ] ) CLOSE_FIX = re.compile(r"(<\|det\|>\s*\w+\s*\[\s*\d+\s*,\s*\d+\s*,\s*\d+\s*,\s*\d+\s*\])\s*<\|?/?det[|/]?>") # <|det|>, <|/det/> ... def norm_box(x0, y0, x1, y1, w, h): f = lambda v, s: max(0, min(999, int(round(v / max(1, s) * 999)))) return f(x0, w), f(y0, h), f(x1, w), f(y1, h) def det(kind, box, content): return f"<|det|>{kind} [{box[0]}, {box[1]}, {box[2]}, {box[3]}]<|/det|>{content}" if eng.startswith("gemini"): KEY = os.environ["GEMINI_API_KEY"] URL = f"https://generativelanguage.googleapis.com/v1beta/models/{eng}:generateContent?key=" elif eng == "mistral": KEY = os.environ["MISTRAL_API_KEY"] elif eng == "clova": CURL, CSEC = os.environ["CLOVA_OCR_URL"], os.environ["CLOVA_OCR_SECRET"] else: raise SystemExit("engine?") def post(url, body, headers): req = urllib.request.Request( url, data=json.dumps(body).encode(), headers={"Content-Type": "application/json", **headers}, ) return json.loads(urllib.request.urlopen(req, timeout=300).read()) def call(r): raw = (root / r["image"].lstrip("/")).read_bytes() im = Image.open(io.BytesIO(raw)) W, H = im.size fmt = ( "webp" if raw[8:12] == b"WEBP" else ("jpeg" if raw[:2] == b"\xff\xd8" else "png") ) if eng.startswith("gemini"): body = { "contents": [ { "parts": [ {"text": PROMPT}, { "inline_data": { "mime_type": f"image/{fmt}", "data": base64.b64encode(raw).decode(), } }, ] } ], "generationConfig": { "temperature": 0, "thinkingConfig": {"thinkingLevel": "minimal"} if eng.startswith("gemini-3.5") else {"thinkingBudget": 0}, "maxOutputTokens": 60000, }, } d = post(URL + KEY, body, {}) c = d["candidates"][0] return ( CLOSE_FIX.sub(r"\1<|/det|>", "".join(p.get("text", "") for p in c.get("content", {}).get("parts", []))), d, c.get("finishReason"), ) if eng == "mistral": body = { "model": "mistral-ocr-latest", "table_format": "html", "include_image_base64": False, "document": { "type": "image_url", "image_url": f"data:image/{fmt};base64," + base64.b64encode(raw).decode(), }, } d = post( "https://api.mistral.ai/v1/ocr", body, {"Authorization": "Bearer " + KEY} ) parts = [] for p in d.get("pages", []): dw, dh = p["dimensions"]["width"], p["dimensions"]["height"] tabs = {t["id"]: t.get("content") or "" for t in p.get("tables") or []} for b in p.get("blocks") or []: box = norm_box( b["top_left_x"], b["top_left_y"], b["bottom_right_x"], b["bottom_right_y"], dw, dh, ) content = ( tabs.get(b.get("table_id") or "", None) if b.get("table_id") else None ) if content is None: content = b.get("content") or "" parts.append(det(b.get("type") or "text", box, content)) return "\n".join(parts), d, "ok" # clova if fmt == "webp" or fmt == "jpeg": b = io.BytesIO() im.convert("RGB").save(b, "PNG") raw = b.getvalue() body = { "version": "V2", "requestId": str(uuid.uuid4()), "timestamp": int(time.time() * 1000), "lang": "ko", "images": [ {"format": "png", "name": "p", "data": base64.b64encode(raw).decode()} ], "enableTableDetection": True, } d = post(CURL, body, {"X-OCR-SECRET": CSEC}) ci = d["images"][0] if ci.get("inferResult") != "SUCCESS": raise RuntimeError(f"inferResult {ci.get('inferResult')} {ci.get('message')}") poly = lambda bp: ( [v.get("x", 0) for v in bp["vertices"]], [v.get("y", 0) for v in bp["vertices"]], ) tables = [] for t in ci.get("tables", []): xs, ys = poly(t["boundingPoly"]) rows = {} for c in t.get("cells", []): txt = " ".join( w["inferText"] for tl in c.get("cellTextLines", []) for w in tl.get("cellWords", []) ) rows.setdefault(c["rowIndex"], []).append( (c["columnIndex"], c.get("columnSpan", 1), c.get("rowSpan", 1), txt) ) h = ( "" + "".join( "" + "".join( " 1 else "") + (f' rowspan="{rs}"' if rs > 1 else "") + f">{html.escape(tx)}" for _, cs, rs, tx in sorted(rows[k]) ) + "" for k in sorted(rows) ) + "
" ) tables.append(((min(xs), min(ys), max(xs), max(ys)), h)) blocks, line, lxs, lys = [], [], [], [] def flush(): if line: x0, y0, x1, y1 = min(lxs), min(lys), max(lxs), max(lys) cx, cy = (x0 + x1) / 2, (y0 + y1) / 2 if not any(a <= cx <= c_ and b <= cy <= d_ for (a, b, c_, d_), _ in tables): blocks.append( ( (y0, x0), det("text", norm_box(x0, y0, x1, y1, W, H), " ".join(line)), ) ) line.clear() lxs.clear() lys.clear() for f in ci.get("fields", []): xs, ys = poly(f["boundingPoly"]) line.append(f["inferText"]) lxs.extend(xs) lys.extend(ys) if f.get("lineBreak"): flush() flush() for (a, b, c_, d_), h in tables: blocks.append(((b, a), det("table", norm_box(a, b, c_, d_, W, H), h))) return "\n".join(x for _, x in sorted(blocks)), d, "SUCCESS" def work(r): t0 = time.time() err = None for a in range(3): try: text, d, fin = call(r) return { "id": r["id"], "text": text, "finish": fin, "error": None, "seconds": round(time.time() - t0, 1), }, d except Exception as e: err = repr(e)[:200] if hasattr(e, "read"): try: err += " " + e.read().decode()[:300] except Exception: pass time.sleep(5 * (a + 1)) return { "id": r["id"], "text": "", "finish": None, "error": err, "seconds": round(time.time() - t0, 1), }, None print(eng, "todo", len(todo), "done", len(done), flush=True) with out.open("a") as f, rawf.open("a") as fr, ThreadPoolExecutor(threads) as ex: for k, (row, d) in enumerate(ex.map(work, todo), 1): f.write(json.dumps(row, ensure_ascii=False) + "\n") f.flush() if d is not None: fr.write( json.dumps({"id": row["id"], "response": d}, ensure_ascii=False) + "\n" ) fr.flush() if k % 50 == 0 or row["error"]: print( k, row["id"], row["error"] or row["finish"], row["seconds"], flush=True ) print("API_DONE", eng, flush=True)