Spaces:
Sleeping
Sleeping
File size: 40,734 Bytes
9381502 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 | # ==============================================
# pdf_parser_adaptive.py (v6.3 CPU-Stable + SmartSkip)
# OCR + LLM ๊ธฐ๋ฐ Adaptive PDF Parser
# โ
๊ธฐ์กด OCR/์ด๋ฏธ์ง ๊ฐ์ง โ OCR ๋จ๊ณ ์๋ ์คํต
# โ
LLM ํ์ฑ๋ง ์ฌ์คํ ๊ฐ๋ฅ
# Compatible with app_v2025 + models_v2025 + llm_parse_v3.1
# ==============================================
import hashlib
import os, re, json, time, traceback
from pathlib import Path
from typing import Dict, List, Tuple, Optional
from dataclasses import dataclass
from concurrent.futures import ThreadPoolExecutor, as_completed
import torch
from pdf2image import convert_from_path
from pdf2image.exceptions import PDFInfoNotInstalledError, PDFPageCountError
from paddleocr import PaddleOCR
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from langchain_huggingface import HuggingFacePipeline
from PIL import Image, ImageOps, ImageEnhance, ImageFilter
# LangChain parser
from cert_study_app.chains.question_parser_chain import parse_question_with_chain
from cert_study_app.services.text_cleanup_service import clean_question_text, is_noise_line
# ----------------------------------------------
# ๊ฒฝ๋ก ์ค์
# ----------------------------------------------
DATA_DIR = Path("data")
IMAGE_DIR = DATA_DIR / "images"
QUESTION_IMAGE_DIR = DATA_DIR / "question_images"
OCR_LOG_DIR = DATA_DIR / "ocr_logs"
RUN_LOG_DIR = DATA_DIR / "run_logs"
for p in [IMAGE_DIR, QUESTION_IMAGE_DIR, OCR_LOG_DIR, RUN_LOG_DIR]:
p.mkdir(parents=True, exist_ok=True)
def _log(level, msg):
print(f"[{level}] {msg}")
def _emit(callback, **payload):
if callback:
callback(payload)
def _job_name(pdf_path: str) -> str:
stem = Path(pdf_path).stem
safe = re.sub(r"[^0-9A-Za-z๊ฐ-ํฃ._-]+", "_", stem).strip("._-") or "pdf"
digest = hashlib.sha1(str(Path(pdf_path).resolve()).encode("utf-8")).hexdigest()[:8]
return f"{safe}_{digest}"
def _item_key(page: int, text: str) -> str:
return hashlib.sha1(f"{page}:{text[:600]}".encode("utf-8")).hexdigest()
def _page_images(directory: Path) -> List[Path]:
paths = sorted(directory.glob("page_*.jpg")) + sorted(directory.glob("page_*.png"))
return [p for p in paths if not p.stem.endswith(("_light", "_heavy"))]
# ----------------------------------------------
# Config
# ----------------------------------------------
@dataclass
class ParserConfig:
pdf_path: str
output_json: str
use_llm: bool = True
lang: str = "korean"
dpi: int = 200
cpu_threads: int = max(1, (os.cpu_count() or 4) // 4)
llm_provider: str = os.getenv("CERT_STUDY_LLM_PROVIDER", "ollama")
llm_model: str = "Qwen/Qwen2.5-1.8B"
ollama_model: str = os.getenv("OLLAMA_MODEL", "qwen2.5:14b")
ollama_base_url: str = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434")
max_new_tokens: int = 192
ocr_workers: int = 1
job_name: str = ""
image_dir: Path = IMAGE_DIR
question_image_dir: Path = QUESTION_IMAGE_DIR
ocr_log_dir: Path = OCR_LOG_DIR
partial_jsonl: Path = RUN_LOG_DIR / "parsed.partial.jsonl"
@dataclass
class PdfTextLine:
page: int
text: str
bbox: Tuple[float, float, float, float]
@dataclass
class QuestionRange:
number: int
start: int
end: int
# ----------------------------------------------
# OCR ์์ฑ๊ธฐ
# ----------------------------------------------
def create_ocr(lang: str, cpu_threads: int) -> PaddleOCR:
return PaddleOCR(
use_angle_cls=True,
lang=lang,
rec_char_type="korean_english",
rec_algorithm="SVTR_LCNet",
det_limit_side_len=1280,
det_db_box_thresh=0.3,
use_gpu=False,
enable_mkldnn=False,
cpu_threads=cpu_threads
)
# ----------------------------------------------
# ์ ์ฒ๋ฆฌ
# ----------------------------------------------
def preprocess_light(img_path: Path) -> Path:
img = Image.open(img_path)
img = ImageOps.exif_transpose(img).convert("L")
img = ImageEnhance.Contrast(img).enhance(1.8)
img = ImageEnhance.Brightness(img).enhance(1.1)
img = img.filter(ImageFilter.MedianFilter(size=3))
out_path = img_path.with_name(img_path.stem + "_light.jpg")
img.save(out_path, "JPEG", quality=90)
return out_path
def preprocess_heavy(img_path: Path) -> Path:
img = Image.open(img_path)
img = ImageOps.exif_transpose(img).convert("L")
w, h = img.size
img = img.resize((int(w * 1.2), int(h * 1.2)))
img = ImageEnhance.Contrast(img).enhance(2.0)
img = ImageEnhance.Sharpness(img).enhance(1.7)
img = img.filter(ImageFilter.GaussianBlur(radius=0.7))
out_path = img_path.with_name(img_path.stem + "_heavy.jpg")
img.save(out_path, "JPEG", quality=90)
return out_path
def ocr_quality_score(text: str) -> float:
if not text.strip():
return 0.0
valid = re.findall(r"[๊ฐ-ํฃA-Za-z0-9]", text)
ratio = len(valid) / max(len(text), 1)
lines = [ln for ln in text.splitlines() if ln.strip()]
avg_len = sum(len(l) for l in lines) / max(1, len(lines))
return min(1.0, ratio * (avg_len / 10 + 0.5))
def clean_text(t: str) -> str:
t = re.sub(r"[\r\t]+", " ", t)
t = re.sub(r"[ \u00A0]+", " ", t)
t = re.sub(r"[ ]{2,}", " ", t)
return t.strip()
def clean_pdf_line(text: str) -> str:
text = clean_text(text)
if not text:
return ""
if is_noise_line(text):
return ""
if "The safer" in text and "IT exams" in text:
return ""
if re.match(r"^\d+\s*/\s*\d+$", text):
return ""
return text
def extract_pdf_text_lines(pdf_path: str) -> List[PdfTextLine]:
try:
import fitz
except Exception:
return []
lines: List[PdfTextLine] = []
try:
with fitz.open(pdf_path) as doc:
for page_index, page in enumerate(doc, 1):
data = page.get_text("dict")
for block in data.get("blocks", []):
if block.get("type") != 0:
continue
for line in block.get("lines", []):
text = "".join(span.get("text", "") for span in line.get("spans", []))
text = clean_pdf_line(text)
if text:
lines.append(
PdfTextLine(
page=page_index,
text=text,
bbox=tuple(line.get("bbox", (0, 0, 0, 0))),
)
)
except Exception as exc:
_log("WARN", f"PDF ํ
์คํธ ์ถ์ถ ์คํจ: {exc}")
return []
return lines
def _question_start_number(text: str) -> Optional[int]:
match = re.match(r"^\s*(\d{1,3})\s*[.)]\s*\S+", text)
if not match:
return None
return int(match.group(1))
def split_pdf_lines_to_ranges(lines: List[PdfTextLine]) -> List[QuestionRange]:
candidates = []
for index, line in enumerate(lines):
number = _question_start_number(line.text)
if number is None:
continue
candidates.append((index, number))
selected = []
expected = 1
for index, number in candidates:
if number == expected:
selected.append((index, number))
expected += 1
ranges: List[QuestionRange] = []
for pos, (start, number) in enumerate(selected):
end = selected[pos + 1][0] if pos + 1 < len(selected) else len(lines)
chunk = "\n".join(line.text for line in lines[start:end])
if "Answer:" not in chunk and "์ ๋ต" not in chunk and len(chunk) < 80:
continue
ranges.append(QuestionRange(number=number, start=start, end=end))
return ranges
def _split_explanation(text: str) -> Tuple[str, str]:
match = re.search(r"(?im)^\s*Explanation\s*:\s*", text)
if not match:
match = re.search(r"(?im)^\s*ํด์ค\s*:\s*", text)
if not match:
return text.strip(), ""
return text[: match.start()].strip(), text[match.end() :].strip()
def _extract_answer(text: str) -> Tuple[str, str]:
match = re.search(r"(?im)^\s*(?:Answer|์ ๋ต)\s*:\s*(.*?)\s*$", text)
if not match:
return "", text
explanation_match = re.search(r"(?im)^\s*(?:Explanation|ํด์ค)\s*:\s*", text[match.end() :])
if explanation_match:
answer_end = match.end() + explanation_match.start()
else:
answer_end = len(text)
answer = (match.group(1) + "\n" + text[match.end() : answer_end]).strip()
answer = re.sub(r"\n+", " ", answer)
answer = re.sub(r"\s+", " ", answer).strip()
without = (text[: match.start()] + "\n" + text[answer_end:]).strip()
return answer, without
def _extract_options(text: str) -> Tuple[str, List[str]]:
option_matches = list(re.finditer(r"(?m)^\s*([A-H])\.\s*(.+)$", text))
if len(option_matches) < 2:
return text.strip(), []
options = []
for pos, match in enumerate(option_matches):
end = option_matches[pos + 1].start() if pos + 1 < len(option_matches) else len(text)
body = text[match.start() : end].strip()
body = re.sub(r"\n+", " ", body)
body = re.sub(r"\s+\d+\s*[.)]\s*[ใใ].*$", "", body)
body = re.sub(r"\s+", " ", body).strip()
options.append(body)
stem = text[: option_matches[0].start()].strip()
return stem, options
def _looks_like_multi_select(text: str, answer: str, options: List[str]) -> bool:
if len(options) < 2:
return False
answer_labels = re.findall(r"\b[A-H]\b", answer or "")
compact_answer = re.sub(r"[^A-H]", "", (answer or "").upper())
has_multiple_answers = len(set(answer_labels)) > 1 or len(set(compact_answer)) > 1
if has_multiple_answers:
return True
return bool(
re.search(
r"(์ด๋ค\s*(๋|์ธ|๋ค)\s*(๊ฐ์ง|๊ฐ)|๋\s*๊ฐ์ง\s*์ต์
|์ธ\s*๊ฐ์ง\s*์์
|๊ฐ\s*์ ๋ต|๊ฐ\s*์ฌ๋ฐ๋ฅธ\s*์ ํ|choose\s+two|choose\s+three|select\s+two|select\s+three)",
text or "",
re.I,
)
)
def _detect_structured_type(text: str, options: List[str], answer: str = "") -> str:
lowered = text.lower()
if "ํซ์คํ" in text or "hotspot" in lowered or "๋ต๋ณ ์์ญ" in text:
return "hotspot"
if "๋์ด" in text or "drag" in lowered or "drop" in lowered:
return "matching"
if "์์" in text or "์ ๋ ฌ" in text or "๋ฐฐ์น" in text:
return "ordering"
if "๋ค์ ๊ฐ ์ง์ " in text or "๊ฐ ์ง์ " in text:
return "yes_no"
if "์" in text and "์๋์ค" in text:
return "yes_no"
if "ํ" in text and ("์ ํ" in text or "๊ฐ ํ" in text):
return "table_choice"
if _looks_like_multi_select(text, answer, options):
return "multi_select"
if len(options) >= 2 and any(marker in text for marker in ["๊ฐ์", "๊ธฐ์กด ํ๊ฒฝ", "์๊ตฌ ์ฌํญ", "๊ณํ๋ ๋ณ๊ฒฝ", "์๋๋ฆฌ์ค"]):
return "case_study"
if len(options) >= 2:
return "mcq"
return "unparsed"
def _group_marker(text: str) -> Optional[Tuple[int, int]]:
match = re.search(r"(\d{1,3})\s*[~๏ฝ-]\s*(\d{1,3})\s*๋ฒ\s*๋ฌธ์ ", text)
if not match:
match = re.search(r"\(\s*(\d{1,3})\s*[~๏ฝ-]\s*(\d{1,3})\s*\)", text)
if not match:
return None
return int(match.group(1)), int(match.group(2))
def _topic_intro_number(text: str) -> Optional[int]:
match = re.match(r"^\s*(\d{1,3})\s*[.)]\s*์ฃผ์ \s*\d+", text)
if not match:
return None
return int(match.group(1))
def _split_parent_and_child(stem: str) -> Tuple[Optional[str], str]:
lines = [line.strip() for line in stem.splitlines() if line.strip()]
if len(lines) < 4:
return None, stem.strip()
marker_index = None
for index, line in enumerate(lines):
if _group_marker(line):
marker_index = index
break
if marker_index is None:
return None, stem.strip()
child_start = max(marker_index + 1, len(lines) - 3)
parent = _clean_parent_stem("\n".join(lines[:child_start]))
child = "\n".join(lines[child_start:]).strip()
return parent, child or stem.strip()
def _split_topic_parent_and_child(stem: str) -> Tuple[Optional[str], str]:
lines = [line.strip() for line in stem.splitlines() if line.strip()]
if len(lines) < 8 or not _topic_intro_number(lines[0]):
return None, stem.strip()
cue_indexes = [
index
for index, line in enumerate(lines)
if any(cue in line for cue in ["ํซ์คํ", "๋๋๊ทธ ์ค ๋๋กญ", "๋๋๊ทธ ๋๋กญ", "๋์ด ๋๊ธฐ"])
]
cue_indexes = [index for index in cue_indexes if index > 2]
if not cue_indexes:
return None, stem.strip()
child_start = cue_indexes[-1]
parent = _clean_parent_stem("\n".join(lines[:child_start]))
child = "\n".join(lines[child_start:]).strip()
return parent, child or stem.strip()
def _clean_parent_stem(text: str) -> str:
cleaned = []
for raw_line in (text or "").splitlines():
line = raw_line.strip()
if not line:
continue
topic_match = re.match(r"^\s*\d{1,3}\s*[.)]\s*(์ฃผ์ \s+\d+\s*,?\s*.+)$", line)
if topic_match:
line = topic_match.group(1).strip()
if re.search(r"\d{1,3}\s*[~๏ฝ-]\s*\d{1,3}\s*๋ฒ\s*๋ฌธ์ \)?", line):
continue
if re.fullmatch(r"\(?\s*\d{1,3}\s*[~๏ฝ-]\s*\d{1,3}\s*\)?", line):
continue
cleaned.append(line)
return "\n".join(cleaned).strip()
def _nonempty_line_count(text: Optional[str]) -> int:
return len([line for line in (text or "").splitlines() if line.strip()])
def _page_image_paths_for_lines(cfg: ParserConfig, lines: List[PdfTextLine], start: int, end: int) -> List[str]:
pages = []
for line in lines[start:end]:
if line.page not in pages:
pages.append(line.page)
paths = []
for page in pages:
path = cfg.image_dir / f"page_{page}.jpg"
if path.exists():
paths.append(path.as_posix())
return paths
def crop_pdf_question_image(cfg: ParserConfig, lines: List[PdfTextLine], qrange: QuestionRange) -> Optional[str]:
cfg.question_image_dir.mkdir(parents=True, exist_ok=True)
start_line = lines[qrange.start]
answer_start = None
for line_index in range(qrange.start, qrange.end):
if re.match(r"(?i)^\s*(Answer|Explanation|Reference)\s*:", lines[line_index].text) or re.match(r"^\s*(์ ๋ต|ํด์ค)\s*:", lines[line_index].text):
answer_start = line_index
break
answer_line = lines[answer_start] if answer_start is not None else None
crop_end = answer_start + 1 if answer_start is not None else qrange.end
answer_lines = [
lines[line_index]
for line_index in range(qrange.start, crop_end)
if re.match(r"(?i)^\s*Answer\s*:", lines[line_index].text) or re.match(r"^\s*์ ๋ต\s*:", lines[line_index].text)
]
end_line = lines[crop_end - 1] if crop_end - 1 < len(lines) else start_line
pages = []
for line in lines[qrange.start : crop_end]:
if line.page not in pages:
pages.append(line.page)
if not pages:
pages = [start_line.page]
crops = []
page_count = len(pages)
scale = cfg.dpi / 72
for page in pages:
page_image = cfg.image_dir / f"page_{page}.jpg"
if not page_image.exists():
continue
with Image.open(page_image) as img:
width, height = img.size
if page == start_line.page:
top = max(0, int(start_line.bbox[1] * scale) - 24)
else:
top = 0
if answer_line and page == answer_line.page:
bottom = max(top + 80, min(height, int(answer_line.bbox[1] * scale) - 12))
elif page == end_line.page:
bottom = min(height, int(end_line.bbox[3] * scale) + 56)
else:
bottom = height
if bottom <= top + 40:
bottom = min(height, top + 240)
crop = img.crop((0, top, width, bottom)).convert("RGB")
page_answer_lines = [line for line in answer_lines if line.page == page]
if page_answer_lines:
from PIL import ImageDraw
draw = ImageDraw.Draw(crop)
for answer_line in page_answer_lines:
mask_top = max(0, int(answer_line.bbox[1] * scale) - top - 8)
mask_bottom = min(crop.height, int(answer_line.bbox[3] * scale) - top + 12)
if mask_bottom > mask_top:
draw.rectangle((0, mask_top, crop.width, mask_bottom), fill=(255, 255, 255))
crops.append(crop.copy())
if not crops:
return None
if len(crops) == 1:
stitched = crops[0]
else:
divider = 18
width = max(crop.width for crop in crops)
height = sum(crop.height for crop in crops) + divider * (len(crops) - 1)
stitched = Image.new("RGB", (width, height), (255, 255, 255))
y = 0
for index, crop in enumerate(crops):
if crop.width != width:
padded = Image.new("RGB", (width, crop.height), (255, 255, 255))
padded.paste(crop, (0, 0))
crop = padded
stitched.paste(crop, (0, y))
y += crop.height
if index < len(crops) - 1:
y += divider
suffix = f"pages_{pages[0]}-{pages[-1]}" if page_count > 1 else f"page_{pages[0]}"
out_path = cfg.question_image_dir / f"q_{qrange.number}_{suffix}.jpg"
stitched.save(out_path, "JPEG", quality=90)
return out_path.as_posix()
def parse_pdf_text_first(cfg: ParserConfig, progress_callback=None) -> List[dict]:
_log("INFO", "[STEP 2] PDF ๋ด์ฅ ํ
์คํธ ๊ธฐ๋ฐ ํ์ฑ ์๋")
lines = extract_pdf_text_lines(cfg.pdf_path)
ranges = split_pdf_lines_to_ranges(lines)
answer_count = sum(1 for line in lines if re.match(r"(?i)^Answer\s*:", line.text))
if len(ranges) < 20 or answer_count < 20:
_log("INFO", f"๋ด์ฅ ํ
์คํธ ๋ฌธํญ ํ๋ณด ๋ถ์กฑ(lines={len(lines)}, ranges={len(ranges)}, answers={answer_count})")
return []
results = []
active_group = None
for index, qrange in enumerate(ranges, 1):
chunk = clean_question_text("\n".join(line.text for line in lines[qrange.start : qrange.end]))
answer, without_answer = _extract_answer(chunk)
before_expl, explanation = _split_explanation(without_answer)
stem, options = _extract_options(before_expl)
question_type = _detect_structured_type(chunk, options, answer)
marker = _group_marker(stem)
parent_stem = None
parent_image_paths = []
group_id = None
topic_intro_number = _topic_intro_number(stem)
if marker:
group_start, group_end = marker
parent_stem, stem = _split_parent_and_child(stem)
parent_line_end = qrange.start + max(_nonempty_line_count(parent_stem), 1)
parent_image_paths = _page_image_paths_for_lines(cfg, lines, qrange.start, min(parent_line_end, qrange.end))
active_group = {
"end": group_end,
"group_id": f"q{group_start}-{group_end}",
"parent_stem": parent_stem or _clean_parent_stem(before_expl),
"parent_image_paths": parent_image_paths,
}
elif topic_intro_number and answer:
parent_stem, split_stem = _split_topic_parent_and_child(stem)
if parent_stem:
stem = split_stem
parent_line_end = qrange.start + max(_nonempty_line_count(parent_stem), 1)
parent_image_paths = _page_image_paths_for_lines(cfg, lines, qrange.start, min(parent_line_end, qrange.end))
active_group = {
"end": None,
"group_id": f"q{topic_intro_number}-case",
"parent_stem": _clean_parent_stem(parent_stem),
"parent_image_paths": parent_image_paths,
}
elif topic_intro_number and not answer and not options:
parent_image_paths = _page_image_paths_for_lines(cfg, lines, qrange.start, qrange.end)
active_group = {
"end": None,
"group_id": f"q{topic_intro_number}-case",
"parent_stem": _clean_parent_stem(before_expl),
"parent_image_paths": parent_image_paths,
}
_emit(
progress_callback,
stage="text_parse",
message=f"๊ณตํต ์ง๋ฌธ ์ ์ฅ q.{topic_intro_number}",
current=index,
total=len(ranges),
)
continue
elif topic_intro_number and active_group and active_group.get("end") is None:
active_group = None
if active_group and active_group.get("end") is None:
group_id = active_group["group_id"]
parent_stem = parent_stem or active_group["parent_stem"]
parent_image_paths = parent_image_paths or active_group.get("parent_image_paths", [])
elif active_group and qrange.number <= active_group["end"]:
group_id = active_group["group_id"]
parent_stem = parent_stem or active_group["parent_stem"]
parent_image_paths = parent_image_paths or active_group.get("parent_image_paths", [])
elif active_group and qrange.number > active_group["end"]:
active_group = None
if marker and not answer and not options:
_emit(
progress_callback,
stage="text_parse",
message=f"๊ณตํต ์ง๋ฌธ ์ ์ฅ q.{qrange.number}",
current=index,
total=len(ranges),
)
continue
image_path = crop_pdf_question_image(cfg, lines, qrange)
result = {
"page": lines[qrange.start].page,
"number": qrange.number,
"group_id": group_id,
"parent_stem": parent_stem,
"parent_image_paths": parent_image_paths,
"stem": stem or before_expl[:1200],
"options": options,
"answer": answer,
"explanation": explanation,
"question_type": question_type,
"image_path": image_path,
"raw_text": chunk,
"parse_status": "draft"
if stem and answer and (options or question_type in {"hotspot", "yes_no", "ordering", "table_choice", "matching"})
else "needs_review",
}
results.append(result)
_emit(
progress_callback,
stage="text_parse",
message=f"PDF ํ
์คํธ ๋ฌธํญ ํ์ฑ {index}/{len(ranges)}",
current=index,
total=len(ranges),
)
_log("INFO", f"[STEP 2] PDF ๋ด์ฅ ํ
์คํธ ํ์ฑ ์๋ฃ ({len(results)}๊ฐ)")
return results
# ----------------------------------------------
# PDF โ ์ด๋ฏธ์ง ๋ณํ
# ----------------------------------------------
def _pdf_page_count(pdf_path: str) -> int:
try:
from pdf2image.pdf2image import pdfinfo_from_path
return int(pdfinfo_from_path(pdf_path).get("Pages") or 0)
except Exception:
import fitz
with fitz.open(pdf_path) as doc:
return doc.page_count
def pdf_to_images(cfg: ParserConfig, progress_callback=None) -> List[Path]:
_log("INFO", f"[STEP 1] PDF โ ์ด๋ฏธ์ง ๋ณํ ์ค... ({cfg.dpi}dpi)")
cfg.image_dir.mkdir(parents=True, exist_ok=True)
try:
page_count = _pdf_page_count(cfg.pdf_path)
except (PDFInfoNotInstalledError, PDFPageCountError) as exc:
_log("WARN", f"PDF ํ์ด์ง ์ ํ์ธ ์คํจ, PyMuPDF๋ก ์ ํํฉ๋๋ค: {exc}")
import fitz
with fitz.open(cfg.pdf_path) as doc:
page_count = doc.page_count
existing = {int(m.group(1)): path for path in _page_images(cfg.image_dir) if (m := re.search(r"page_(\d+)", path.name))}
if page_count and len(existing) >= page_count:
_log("INFO", f"[STEP 1] ๊ธฐ์กด ์ด๋ฏธ์ง {len(existing)}๊ฐ ๋ฐ๊ฒฌ โ ๋ณํ ์คํต")
_emit(
progress_callback,
stage="render",
message=f"๊ธฐ์กด ํ์ด์ง ์ด๋ฏธ์ง {len(existing)}๊ฐ๋ฅผ ์ฌ์ฉํฉ๋๋ค.",
current=len(existing),
total=page_count,
)
return [existing[i] for i in sorted(existing)]
paths = []
for page_num in range(1, page_count + 1):
path = cfg.image_dir / f"page_{page_num}.jpg"
if path.exists():
paths.append(path)
_emit(
progress_callback,
stage="render",
message=f"๊ธฐ์กด ํ์ด์ง ์ด๋ฏธ์ง ์ฌ์ฉ p.{page_num}",
current=page_num,
total=page_count,
)
continue
try:
pages = convert_from_path(
cfg.pdf_path,
dpi=cfg.dpi,
first_page=page_num,
last_page=page_num,
thread_count=1,
)
if pages:
pages[0].save(path, "JPEG", quality=90)
paths.append(path)
except PDFInfoNotInstalledError:
_log("WARN", "Poppler๋ฅผ ์ฐพ์ ์ ์์ด PyMuPDF ๋ ๋๋ง์ผ๋ก ์ ํํฉ๋๋ค.")
import fitz
zoom = cfg.dpi / 72
matrix = fitz.Matrix(zoom, zoom)
with fitz.open(cfg.pdf_path) as doc:
page = doc.load_page(page_num - 1)
pix = page.get_pixmap(matrix=matrix, alpha=False)
pix.save(path)
paths.append(path)
_emit(
progress_callback,
stage="render",
message=f"ํ์ด์ง ์ด๋ฏธ์ง ๋ณํ p.{page_num}",
current=page_num,
total=page_count,
)
_log("INFO", f"[STEP 1] ์๋ฃ ({len(paths)}ํ์ด์ง)")
return paths
# ----------------------------------------------
# OCR ์คํ (์๋ ์คํต)
# ----------------------------------------------
def run_ocr(cfg: ParserConfig, ocr: PaddleOCR, img_paths: List[Path]) -> Dict[int, str]:
_log("INFO", "[STEP 2] OCR ๋จ๊ณ ์์ (์๋ ์คํต ๊ฐ์ง)")
ocr_map = {}
cfg.ocr_log_dir.mkdir(parents=True, exist_ok=True)
existing_txts = sorted(cfg.ocr_log_dir.glob("page_*.txt"))
# โ
๊ธฐ์กด OCR ํ
์คํธ ๊ฒฐ๊ณผ ์กด์ฌ ์์๋ง ์คํต
if existing_txts:
_log("INFO", f"๊ธฐ์กด OCR ๊ฒฐ๊ณผ ๊ฐ์ง๋จ โ OCR ๋จ๊ณ ์คํตํ๊ณ ๊ธฐ์กด ๊ฒฐ๊ณผ ์ฌ์ฉ")
# ํ
์คํธ ํ์ผ ์ฐ์ ๋ก๋
for txt_path in existing_txts:
m = re.search(r"page[_-]?(\d+)", txt_path.name)
page_num = int(m.group(1)) if m else 0
try:
with open(txt_path, "r", encoding="utf-8") as f:
text = clean_text(f.read())
ocr_map[page_num] = text
except Exception as e:
_log("WARN", f"OCR ํ
์คํธ ๋ก๋ ์คํจ: {txt_path} ({e})")
return ocr_map
# โ
OCR ์ ๊ท ์ํ
_log("INFO", "[STEP 2] OCR ์ ๊ท ์ํ ์ค...")
def ocr_one(img_path: Path):
m = re.search(r"page[_-]?(\d+)", img_path.name)
page = int(m.group(1)) if m else 0
cache_path = cfg.ocr_log_dir / f"page_{page}.txt"
for pre_fn in [preprocess_light, preprocess_heavy]:
try:
processed = pre_fn(img_path)
res = ocr.ocr(str(processed), cls=True)
text = "\n".join([ln[1][0] for ln in (res[0] or [])]) if res and res[0] else ""
text = clean_text(text)
if text and ocr_quality_score(text) >= 0.3:
cache_path.write_text(text, encoding="utf-8")
return page, text
except Exception as e:
_log("WARN", f"OCR ์คํจ (p{page}): {e}")
cache_path.write_text("", encoding="utf-8")
return page, ""
with ThreadPoolExecutor(max_workers=cfg.ocr_workers) as ex:
futures = [ex.submit(ocr_one, p) for p in img_paths]
for i, f in enumerate(as_completed(futures), 1):
p, txt = f.result()
ocr_map[p] = txt
if i % 2 == 0:
_log("INFO", f"[OCR] {i}/{len(img_paths)} ์๋ฃ (p{p}, score={ocr_quality_score(txt):.2f})")
return ocr_map
# ----------------------------------------------
# ๋ณํฉ ๋ฐ LLM ํ์ฑ
# ----------------------------------------------
Q_SPLIT = re.compile(r"(?:๋ฌธ์ \s*\d+\.?|Q\s*\d+\.?|^\s*\d+\.\s|๋ค์\s*์ค|์๋๋ฆฌ์ค|Case\s*\d+|Explanation\s*:?)", re.IGNORECASE | re.MULTILINE)
def merge_pages_to_questions(ocr_map: Dict[int, str]) -> List[Tuple[int, str]]:
merged = []
for p in sorted(ocr_map.keys()):
txt = clean_text(ocr_map[p])
if not txt:
continue
chunks = [c.strip() for c in re.split(Q_SPLIT, txt) if c.strip()]
for c in chunks:
if len(c) >= 50 and any(k in c for k in ["์ ๋ต", "๋ณด๊ธฐ", "Answer", "Explanation", "๋ฌธ์ "]):
merged.append((p, c))
if not merged:
for p in sorted(ocr_map.keys()):
txt = clean_text(ocr_map[p])
if len(txt) >= 50 and not txt.startswith("[SKIP OCR]"):
merged.append((p, txt))
return merged
def load_partial_results(partial_jsonl: Path) -> Tuple[List[dict], set]:
results = []
keys = set()
if not partial_jsonl.exists():
return results, keys
with partial_jsonl.open("r", encoding="utf-8") as f:
for line in f:
try:
row = json.loads(line)
except Exception:
continue
item = row.get("result") if isinstance(row, dict) else None
key = row.get("key") if isinstance(row, dict) else None
if item and key:
results.append(item)
keys.add(key)
return results, keys
def append_partial_result(partial_jsonl: Path, key: str, result: dict) -> None:
partial_jsonl.parent.mkdir(parents=True, exist_ok=True)
with partial_jsonl.open("a", encoding="utf-8") as f:
f.write(json.dumps({"key": key, "result": result}, ensure_ascii=False) + "\n")
def create_question_crops(
cfg: ParserConfig,
items: List[Tuple[int, str]],
progress_callback=None,
) -> Dict[int, str]:
cfg.question_image_dir.mkdir(parents=True, exist_ok=True)
by_page: Dict[int, List[int]] = {}
for idx, (page, _text) in enumerate(items):
by_page.setdefault(page, []).append(idx)
image_paths: Dict[int, str] = {}
total = max(len(items), 1)
done = 0
for page, item_indexes in by_page.items():
page_image = cfg.image_dir / f"page_{page}.jpg"
if not page_image.exists():
continue
with Image.open(page_image) as img:
width, height = img.size
band_count = len(item_indexes)
for position, item_index in enumerate(item_indexes):
top = max(0, int(height * position / band_count) - 60)
bottom = min(height, int(height * (position + 1) / band_count) + 60)
crop = img.crop((0, top, width, bottom))
out_path = cfg.question_image_dir / f"page_{page}_q{position + 1}.jpg"
crop.save(out_path, "JPEG", quality=88)
image_paths[item_index] = out_path.as_posix()
done += 1
_emit(
progress_callback,
stage="crop",
message=f"๋ฌธํญ ์ด๋ฏธ์ง ์์ฑ p.{page} ({done}/{total})",
current=done,
total=total,
)
return image_paths
def load_llm(cfg: ParserConfig):
if cfg.llm_provider == "ollama":
_log("INFO", f"[STEP 4] Ollama LLM ์ฐ๊ฒฐ ์ค... ({cfg.ollama_model}, {cfg.ollama_base_url})")
from langchain_community.llms import Ollama
return Ollama(
model=cfg.ollama_model,
base_url=cfg.ollama_base_url,
temperature=0,
)
_log("INFO", f"[STEP 4] Hugging Face LLM ๋ก๋ ์ค... ({cfg.llm_model})")
tok = AutoTokenizer.from_pretrained(cfg.llm_model)
mdl = AutoModelForCausalLM.from_pretrained(cfg.llm_model, torch_dtype="auto", low_cpu_mem_usage=True)
pipe = pipeline("text-generation", model=mdl, tokenizer=tok, device=-1, max_new_tokens=cfg.max_new_tokens, temperature=0.0)
_log("INFO", "[STEP 4] LLM ๋ก๋ ์๋ฃ")
return HuggingFacePipeline(pipeline=pipe)
# ----------------------------------------------
# ๋ฉ์ธ ํ์ดํ๋ผ์ธ
# ----------------------------------------------
def parse_pdf(
pdf_path: str,
output_json: str,
use_llm=True,
lang="korean",
dpi: Optional[int] = None,
llm_provider: Optional[str] = None,
llm_model: Optional[str] = None,
ollama_base_url: Optional[str] = None,
progress_callback=None,
):
torch.set_num_threads(1)
cfg = ParserConfig(pdf_path, output_json, use_llm, lang)
cfg.job_name = _job_name(pdf_path)
cfg.image_dir = IMAGE_DIR / cfg.job_name
cfg.question_image_dir = QUESTION_IMAGE_DIR / cfg.job_name
cfg.ocr_log_dir = OCR_LOG_DIR / cfg.job_name
cfg.partial_jsonl = RUN_LOG_DIR / f"{cfg.job_name}.partial.jsonl"
if dpi:
cfg.dpi = dpi
if llm_provider:
cfg.llm_provider = llm_provider
if llm_model:
if cfg.llm_provider == "ollama":
cfg.ollama_model = llm_model
else:
cfg.llm_model = llm_model
if ollama_base_url:
cfg.ollama_base_url = ollama_base_url
_log("INFO", f"[START] {pdf_path} (dpi={cfg.dpi}, threads={cfg.cpu_threads})")
_emit(progress_callback, stage="start", message="ํ์ฑ์ ์์ํฉ๋๋ค.", current=0, total=1)
# STEP 1: PDF โ ์ด๋ฏธ์ง (์์ผ๋ฉด ์คํต)
_emit(progress_callback, stage="render", message="PDF๋ฅผ ํ์ด์ง ์ด๋ฏธ์ง๋ก ๋ณํํฉ๋๋ค.", current=0, total=1)
pages = pdf_to_images(cfg, progress_callback=progress_callback)
_emit(progress_callback, stage="render", message=f"ํ์ด์ง ์ด๋ฏธ์ง {len(pages)}๊ฐ ์ค๋น ์๋ฃ", current=len(pages), total=len(pages) or 1)
text_results = parse_pdf_text_first(cfg, progress_callback=progress_callback)
if text_results:
Path(output_json).write_text(json.dumps(text_results, ensure_ascii=False, indent=2), encoding="utf-8")
_log("INFO", f"[DONE] PDF ํ
์คํธ ๊ธฐ๋ฐ ์ด {len(text_results)} ๋ฌธํญ ์ ์ฅ โ {output_json}")
_emit(
progress_callback,
stage="done",
message=f"PDF ํ
์คํธ ๊ธฐ๋ฐ {len(text_results)}๋ฌธํญ ํ์ฑ ์๋ฃ",
current=len(text_results),
total=len(text_results),
)
return text_results
# STEP 2: OCR (์์ผ๋ฉด ์คํต)
_emit(progress_callback, stage="ocr", message="OCR ํ
์คํธ๋ฅผ ์ค๋นํฉ๋๋ค.", current=0, total=len(pages) or 1)
ocr = create_ocr(cfg.lang, cfg.cpu_threads)
ocr_map = run_ocr(cfg, ocr, pages)
_emit(progress_callback, stage="ocr", message=f"OCR ํ
์คํธ {len(ocr_map)}ํ์ด์ง ์ค๋น ์๋ฃ", current=len(ocr_map), total=len(pages) or len(ocr_map) or 1)
# STEP 3: ๋ฌธํญ ๋ณํฉ
items = merge_pages_to_questions(ocr_map)
_log("INFO", f"[STEP 3] ๋ฌธํญ ๋ณํฉ ์๋ฃ ({len(items)}๊ฐ)")
_emit(progress_callback, stage="split", message=f"๋ฌธํญ ํ๋ณด {len(items)}๊ฐ๋ฅผ ์ฐพ์์ต๋๋ค.", current=0, total=len(items) or 1)
item_images = create_question_crops(cfg, items, progress_callback=progress_callback)
# STEP 4: LLM ํ์ฑ
results, parsed_keys = load_partial_results(cfg.partial_jsonl)
if results:
_log("INFO", f"[RESUME] ๋ถ๋ถ ์ ์ฅ ๊ฒฐ๊ณผ {len(results)}๊ฐ ๋ก๋ โ ์ด๋ฏธ ์ฒ๋ฆฌํ ๋ฌธํญ ์คํต")
_emit(
progress_callback,
stage="resume",
message=f"์ด์ ๋ถ๋ถ ์ ์ฅ {len(results)}๊ฐ๋ฅผ ์ด์ด์ ์ฌ์ฉํฉ๋๋ค.",
current=len(results),
total=len(items) or len(results) or 1,
)
if use_llm:
try:
llm = load_llm(cfg)
except Exception as e:
_log("WARN", f"LLM ๋ก๋ ์คํจ โ ๊ธฐ๋ณธ ํ์ฑ์ผ๋ก ๊ณ์ ์งํํฉ๋๋ค: {e}")
use_llm = False
else:
for index, (p, t) in enumerate(items, 1):
key = _item_key(p, t)
if key in parsed_keys:
_emit(
progress_callback,
stage="llm",
message=f"์ด๋ฏธ ์ฒ๋ฆฌํ ๋ฌธํญ์ ๊ฑด๋๋๋๋ค. ({index}/{len(items)})",
current=index,
total=len(items),
)
continue
_emit(
progress_callback,
stage="llm",
message=f"Qwen์ผ๋ก ๋ฌธํญ์ ์ ์ ํฉ๋๋ค. p.{p} ({index}/{len(items)})",
current=index - 1,
total=len(items),
)
try:
parsed = parse_question_with_chain(llm, p, t)
except Exception as e:
parsed = {
"page": p,
"stem": t[:400],
"options": [],
"answer": [],
"explanation": f"LLM ์คํจ: {e}",
"question_type": "mcq"
}
parsed["raw_text"] = parsed.get("raw_text") or t
parsed["chunk_index"] = index
parsed["chunk_key"] = key
if index - 1 in item_images:
parsed["image_path"] = item_images[index - 1]
results.append(parsed)
parsed_keys.add(key)
append_partial_result(cfg.partial_jsonl, key, parsed)
Path(output_json).write_text(json.dumps(results, ensure_ascii=False, indent=2), encoding="utf-8")
_emit(
progress_callback,
stage="llm",
message=f"๋ฌธํญ {len(results)}๊ฐ ์ ์ฅ๋จ",
current=index,
total=len(items),
)
if not use_llm:
for index, (p, t) in enumerate(items, 1):
key = _item_key(p, t)
if key in parsed_keys:
continue
parsed = {
"page": p,
"stem": t[:400],
"options": [],
"answer": [],
"explanation": "",
"question_type": "mcq",
"image_path": item_images.get(index - 1),
"raw_text": t,
"chunk_index": index,
"chunk_key": key,
}
results.append(parsed)
parsed_keys.add(key)
append_partial_result(cfg.partial_jsonl, key, parsed)
_emit(progress_callback, stage="basic", message=f"๊ธฐ๋ณธ ํ์ฑ {index}/{len(items)}", current=index, total=len(items))
if not use_llm and not results:
for index, (p, t) in enumerate(items, 1):
results.append({
"page": p,
"stem": t[:400],
"options": [],
"answer": [],
"explanation": "",
"question_type": "mcq",
"raw_text": t,
"chunk_index": index,
"chunk_key": _item_key(p, t),
})
Path(output_json).write_text(json.dumps(results, ensure_ascii=False, indent=2), encoding="utf-8")
_log("INFO", f"[DONE] ์ด {len(results)} ๋ฌธํญ ์ ์ฅ โ {output_json}")
_emit(progress_callback, stage="done", message=f"์ด {len(results)}๋ฌธํญ ํ์ฑ ์๋ฃ", current=len(results), total=len(results) or 1)
return results
# ----------------------------------------------
# CLI ๋๋ฒ๊ทธ ์คํ
# ----------------------------------------------
if __name__ == "__main__":
import argparse
ap = argparse.ArgumentParser()
ap.add_argument("--pdf", required=True)
ap.add_argument("--out", default="data/parsed.json")
ap.add_argument("--no-llm", action="store_true")
ap.add_argument("--dpi", type=int, default=None)
args = ap.parse_args()
try:
parse_pdf(args.pdf, args.out, use_llm=not args.no_llm, dpi=args.dpi)
except Exception as e:
_log("ERROR", f"์คํ ์ค ์ค๋ฅ: {e}")
print(traceback.format_exc())
|