Spaces:
Sleeping
Sleeping
File size: 40,005 Bytes
9f2df60 3d053f1 9f2df60 7435859 9f2df60 7435859 9f2df60 7435859 9f2df60 34406a0 3d053f1 34406a0 9f2df60 34406a0 9f2df60 0096fae 9f2df60 3d053f1 9f2df60 3d053f1 9f2df60 3d053f1 34406a0 9f2df60 7435859 9f2df60 7435859 9f2df60 0096fae 9f2df60 0096fae 9f2df60 7435859 9f2df60 7435859 9f2df60 7435859 9f2df60 7435859 3d053f1 7435859 3d053f1 7435859 9f2df60 7435859 3d053f1 7435859 9f2df60 7435859 9f2df60 7435859 9f2df60 | 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 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 | """
Regenerate data/database/programme_facts.json from the official programme websites.
Offline fact-extraction step (multi-agent offline, single-agent online):
this script runs OUTSIDE the chat request path — manually, via cron, or as a
post-scrape pipeline step. It fetches the official sources, lets an LLM
extract the volatile core facts into a strict schema, diffs against the
current facts file, and alerts via the notification center when facts changed.
Usage:
python -m src.pipeline.update_programme_facts # update + diff alert
python -m src.pipeline.update_programme_facts --dry-run # show diff only
"""
import argparse
import html
import json
import os
import re
import sys
import unicodedata
from datetime import date
from tempfile import NamedTemporaryFile
import requests
from pydantic import BaseModel, Field
from src.config import config
from src.utils.logging import get_logger
logger = get_logger('update_programme_facts')
FACTS_PATH = os.path.join(config.paths.DATA, 'database', 'programme_facts.json')
# Pages and data-plan PDFs that contain the volatile core facts.
FACT_SOURCES = {
'overview': 'https://emba.unisg.ch/',
'deadlines': 'https://emba.unisg.ch/bewerbung/fristen',
'emba': 'https://emba.unisg.ch/programm/emba',
'iemba': 'https://emba.unisg.ch/programm/iemba',
'iemba_es': 'https://es.unisg.ch/en/executive-programme/international-executive-mba-hsg/',
'emba_x': 'https://embax.ch/',
'emba_plan': 'https://emba.unisg.ch/wp-content/uploads/2026/05/Neuer-Dataplan-EMBA71-mitRatenplan.pdf',
'iemba_plan': 'https://emba.unisg.ch/wp-content/uploads/2026/05/IEMBA-14-info-sheet-with-payment-plan-6.pdf',
}
REQUEST_TIMEOUT = 30
REQUEST_HEADERS = {
'User-Agent': (
'Mozilla/5.0 (Windows NT 10.0; Win64; x64) '
'AppleWebKit/537.36 (KHTML, like Gecko) '
'Chrome/125.0 Safari/537.36'
),
'Accept': 'text/html,application/xhtml+xml,application/xml;q=0.9,application/pdf;q=0.9,*/*;q=0.8',
'Accept-Language': 'de-CH,de;q=0.9,en;q=0.8',
'Accept-Encoding': 'gzip, deflate',
'Cache-Control': 'no-cache',
}
FALLBACK_REQUEST_HEADERS = {
**REQUEST_HEADERS,
'Referer': 'https://emba.unisg.ch/',
}
ACCESS_CHALLENGE_MARKERS = (
'please wait while your request is being verified',
'checking your browser before accessing',
'enable javascript and cookies to continue',
'verify you are human',
)
# ----------------------------- Extraction schema -----------------------------
class DeadlineFee(BaseModel):
deadline: str = Field(description="Application deadline as ISO date YYYY-MM-DD")
fee: int = Field(description="Tuition fee in CHF as plain integer, e.g. 77500")
class BilingualText(BaseModel):
de: str = Field(description="German wording")
en: str = Field(description="English wording")
class ProgrammeFactsSchema(BaseModel):
official_name: str
current_cohort: str = Field(description="e.g. 'EMBA 71', 'IEMBA 14', 'emba X6'")
language: BilingualText = Field(description="Programme teaching language")
programme_start: str = Field(description="ISO date YYYY-MM-DD of the next cohort start")
duration: BilingualText
ects_credits: int = Field(default=0, description="ECTS credits as plain integer, e.g. 75; 0 if missing")
structure: BilingualText = Field(description="Courses, campus weeks, projects")
locations: BilingualText
first_deadline: DeadlineFee
final_deadline: DeadlineFee
advisor_name: str
advisor_email: str
advisor_phone: str
class AllProgrammesSchema(BaseModel):
emba: ProgrammeFactsSchema
iemba: ProgrammeFactsSchema
emba_x: ProgrammeFactsSchema
class FactComparisonDecision(BaseModel):
materially_changed: bool
confidence: float = Field(ge=0.0, le=1.0)
reason: str
fact_value: str
preserve_existing: bool
EXTRACTION_PROMPT = """You are a fact extraction system. Below is the text content of the
official HSG Executive MBA websites. Extract the CURRENT facts for the three
programmes EMBA HSG (German), IEMBA HSG (International, English) and
emba X (ETH Zurich & University of St.Gallen joint degree, English).
Rules:
- Use ONLY facts that literally appear in the provided page content.
- Never guess or fill gaps from prior knowledge. If a value is genuinely
missing from the pages, use an empty string.
- Fees are CHF integers without separators (CHF 77'500 -> 77500).
- ECTS credits are plain integers (75 ECTS -> 75). If missing, use 0.
- Dates in ISO format (14. September 2026 -> 2026-09-14).
- Never mix values between programmes. The deadlines page contains one row
per programme - keep them strictly separated.
- Currently stored facts are provided for stability and comparison only. Do not
use them to fill missing page content, but if the page expresses the same
fact with different punctuation, word order, translation-equivalent wording,
or minor synonyms, prefer the existing stable wording.
CURRENTLY STORED FACTS:
{existing_facts_context}
PAGE CONTENT:
{page_content}"""
FACT_COMPARISON_PROMPT = """You compare one stored programme fact with a newly
observed fact extracted from official page content.
Rules:
- Return materially_changed=false when the page expresses the same factual
content, even if wording, punctuation, formatting, translation, or synonyms
differ.
- Return materially_changed=true only for real factual differences: fees,
deadlines, start dates, numbers of courses/modules/electives, campus weeks,
admissions requirements, duration, degree/certificate/title, language,
location, format, or a component being added or removed.
- Be conservative. If the difference is stylistic or ambiguous, preserve the
existing value and set preserve_existing=true.
- If the page contains the same information expressed differently, keep the
existing stored fact as fact_value.
Fact key: {fact_key}
Language: {language}
Source: {source_info}
Currently stored value:
{existing_value}
Newly observed/extracted value:
{observed_value}
Relevant page snippet:
{page_content}"""
# --------------------------------- Fetching ----------------------------------
def extract_pdf_text(content: bytes, url: str) -> str:
"""Extract text from a PDF response using available local parsers."""
if not content.lstrip().startswith(b'%PDF'):
logger.warning(f"PDF URL did not return PDF bytes: {url}")
return ''
suffix = os.path.splitext(url)[1] or '.pdf'
with NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
tmp.write(content)
tmp_path = tmp.name
try:
try:
from docling.document_converter import DocumentConverter
result = DocumentConverter().convert(tmp_path)
return result.document.export_to_markdown()
except Exception as docling_error:
logger.warning(f"Docling could not parse PDF {url}; trying fallback parser: {docling_error}")
try:
from pypdf import PdfReader
reader = PdfReader(tmp_path)
return "\n\n".join(page.extract_text() or '' for page in reader.pages).strip()
except Exception as pypdf_error:
logger.warning(f"Fallback PDF parser could not parse {url}: {pypdf_error}")
raise
finally:
try:
os.remove(tmp_path)
except OSError:
pass
def _extract_fact_html_snippets(text: str) -> str:
"""Keep structured fact blocks before converting the page to visible text."""
matches = re.findall(
r'<div[^>]*class=["\'][^"\']*\blocations\b[^"\']*["\'][^>]*>.*?</div>',
text or '',
flags=re.IGNORECASE | re.DOTALL,
)
matches += [
table
for table in re.findall(
r'<table[^>]*>.*?</table>', text or '', flags=re.IGNORECASE | re.DOTALL
)
if re.search(r'\battendance\b|Pflichtkurse', table, flags=re.IGNORECASE)
]
return "\n".join(matches)
def _is_access_challenge(text: str) -> bool:
normalized = (text or '').casefold()
return any(marker in normalized for marker in ACCESS_CHALLENGE_MARKERS)
def fetch_sources() -> dict[str, str]:
"""Fetch all fact source pages. Raises when a page cannot be fetched."""
pages = {}
session = requests.Session()
for key, url in FACT_SOURCES.items():
logger.info(f"Fetching {url}")
resp = session.get(url, timeout=REQUEST_TIMEOUT, headers=REQUEST_HEADERS)
if resp.status_code == 415:
logger.warning(f"Retrying {url} after HTTP 415 with fallback headers")
resp = session.get(url, timeout=REQUEST_TIMEOUT, headers=FALLBACK_REQUEST_HEADERS)
resp.raise_for_status()
content_type = resp.headers.get('Content-Type', '').lower()
if _is_access_challenge(resp.text):
logger.warning(
"Skipping access-challenge response for %s (status=%s, content-type=%s, final-url=%s)",
url,
resp.status_code,
content_type or '<missing>',
getattr(resp, 'url', url),
)
pages[key] = ''
continue
if url.lower().endswith('.pdf') or 'application/pdf' in content_type:
if not resp.content.lstrip().startswith(b'%PDF'):
logger.warning(
"Skipping non-PDF response for %s (status=%s, content-type=%s, final-url=%s)",
url,
resp.status_code,
content_type or '<missing>',
getattr(resp, 'url', url),
)
pages[key] = ''
continue
try:
pages[key] = extract_pdf_text(resp.content, url)
except Exception as exc:
logger.warning(f"Skipping unreadable PDF source {url}: {exc}")
pages[key] = ''
continue
# Lightweight HTML -> text. The scraping pipeline has richer
# processors; for fact extraction visible text is sufficient.
fact_html = _extract_fact_html_snippets(resp.text)
try:
from bs4 import BeautifulSoup
soup = BeautifulSoup(resp.text, 'html.parser')
for tag in soup(['script', 'style', 'noscript']):
tag.decompose()
visible_text = soup.get_text(separator='\n', strip=True)
pages[key] = "\n\n".join(part for part in (fact_html, visible_text) if part)
except ImportError:
pages[key] = resp.text
return pages
# -------------------------------- Extraction ---------------------------------
def _existing_facts_context(existing_facts: dict | None) -> str:
if not existing_facts:
return "No currently stored facts were provided."
return json.dumps(
existing_facts.get('programmes', existing_facts),
indent=2,
ensure_ascii=False,
)[:20000]
def extract_facts(pages: dict[str, str], existing_facts: dict | None = None) -> AllProgrammesSchema:
"""LLM-based structured extraction over the fetched pages."""
from src.rag.models import ModelConfigurator
model = ModelConfigurator.get_main_agent_model().with_structured_output(
AllProgrammesSchema
)
page_content = "\n\n".join(
f"===== SOURCE: {FACT_SOURCES[key]} =====\n{text[:20000]}"
for key, text in pages.items()
)
return model.invoke(EXTRACTION_PROMPT.format(
existing_facts_context=_existing_facts_context(existing_facts),
page_content=page_content,
))
def _extract_ects_credits(text: str) -> int:
"""Deterministically extract ECTS credits from nearby label/value text."""
patterns = [
r'ECTS[-\s]*(?:Punkte|Credits?)\s*[:\n\r\s]+(\d{1,3})\b',
r'(\d{1,3})\s*(?:ECTS|Credits?)\b',
]
for pattern in patterns:
match = re.search(pattern, text, flags=re.IGNORECASE)
if match:
return int(match.group(1))
return 0
def apply_deterministic_fallbacks(extracted: AllProgrammesSchema, pages: dict[str, str]) -> AllProgrammesSchema:
"""Fill simple numeric facts that the LLM occasionally misses."""
fallback_sources = {
'emba': ['emba_plan', 'emba'],
'iemba': ['iemba_es', 'iemba_plan', 'iemba'],
'emba_x': ['emba_x'],
}
for programme_key, source_keys in fallback_sources.items():
programme = getattr(extracted, programme_key)
if programme.ects_credits:
continue
for source_key in source_keys:
ects = _extract_ects_credits(pages.get(source_key, ''))
if ects:
programme.ects_credits = ects
break
return extracted
LOCATION_TRANSLATIONS = {
'Belgien': 'Belgium',
'Belgium': 'Belgium',
'Beijing': 'Beijing',
'China': 'China',
'Costa Rica': 'Costa Rica',
'Italien': 'Italy',
'Italy': 'Italy',
'Japan': 'Japan',
'Peking': 'Beijing',
'Schweiz': 'Switzerland',
'Switzerland': 'Switzerland',
'South Africa': 'South Africa',
'Spanien': 'Spain',
'Spain': 'Spain',
'Südafrika': 'South Africa',
'Tokyo': 'Tokyo',
'Tokio': 'Tokyo',
'USA': 'USA',
}
LOCATION_COUNTRIES_DE = set(LOCATION_TRANSLATIONS)
LOCATION_ELECTIVE_MARKERS = {'wahlkurs', 'elective course', 'elective'}
LOCATION_SECTION_STARTS = {'orte', 'locations'}
LOCATION_SECTION_ENDS = {
'courses',
'course structure',
'duration',
'fees',
'programme structure',
'start',
'total',
'kurse',
'dauer',
'gebühr',
'programmstruktur',
'start',
}
def _clean_html_fragment(value: str) -> str:
value = re.sub(r'<[^>]+>', '', value)
value = html.unescape(value)
return re.sub(r'\s+', ' ', value).strip()
def _canonicalize_location_de(value: str) -> str:
value = re.sub(r'\s+', ' ', value).strip()
parts = [part.strip() for part in value.split(',')]
if len(parts) == 2 and parts[1] in LOCATION_COUNTRIES_DE:
return f"{parts[1]} ({parts[0]})"
return value
def _translate_location_name(value: str) -> str:
match = re.fullmatch(r'(.+?) \((.+)\)', value)
if match:
country_de, place_de = match.groups()
country_en = LOCATION_TRANSLATIONS.get(country_de, country_de)
place_en = LOCATION_TRANSLATIONS.get(place_de, place_de)
return f"{country_en} ({place_en})"
return LOCATION_TRANSLATIONS.get(value, value)
def _locations_from_items(items: list[tuple[str, bool]]) -> BilingualText | None:
de_locations = []
en_locations = []
for location_de, is_elective in items:
location_de = _canonicalize_location_de(location_de)
if not location_de:
continue
location_en = _translate_location_name(location_de)
if is_elective:
location_de = f"{location_de} (Wahlkurs)"
location_en = f"{location_en} (elective)"
de_locations.append(location_de)
en_locations.append(location_en)
if not de_locations:
return None
return BilingualText(de=', '.join(de_locations), en=', '.join(en_locations))
def _extract_locations_from_html(text: str) -> BilingualText | None:
match = re.search(
r'<div[^>]*class=["\'][^"\']*\blocations\b[^"\']*["\'][^>]*>\s*'
r'<small>\s*Orte\s*</small>\s*<ul[^>]*>(.*?)</ul>',
text or '',
flags=re.IGNORECASE | re.DOTALL,
)
if not match:
return None
items = []
for item_html in re.findall(r'<li>(.*?)</li>', match.group(1), flags=re.IGNORECASE | re.DOTALL):
is_elective = re.search(r'<small[^>]*>\s*Wahlkurs\s*</small>', item_html, flags=re.IGNORECASE)
location_de = _clean_html_fragment(
re.sub(r'<small[^>]*>.*?</small>', '', item_html, flags=re.IGNORECASE | re.DOTALL)
)
if location_de:
items.append((location_de, bool(is_elective)))
return _locations_from_items(items)
def _extract_locations_from_text(text: str) -> BilingualText | None:
lines = [
_clean_html_fragment(line)
for line in (text or '').splitlines()
if _clean_html_fragment(line)
]
start_index = None
for index, line in enumerate(lines):
if _canonical_text(line) in LOCATION_SECTION_STARTS:
start_index = index + 1
break
if start_index is None:
return None
items = []
index = start_index
while index < len(lines):
line = lines[index]
canonical_line = _canonical_text(line)
if canonical_line in LOCATION_SECTION_ENDS:
break
if canonical_line in LOCATION_ELECTIVE_MARKERS:
index += 1
continue
next_line = lines[index + 1] if index + 1 < len(lines) else ''
is_elective = _canonical_text(next_line) in LOCATION_ELECTIVE_MARKERS
items.append((line, is_elective))
index += 2 if is_elective else 1
return _locations_from_items(items)
def _extract_locations_from_programme_page(text: str) -> BilingualText | None:
"""Deterministically parse the official programme-page locations block."""
return _extract_locations_from_html(text) or _extract_locations_from_text(text)
STRUCTURE_EXTRA_TRANSLATIONS = {
'Diplomarbeit': 'thesis',
'Capstone-Projekt': 'capstone project',
'Selbststudium': 'self-study',
}
def _extract_structure_from_programme_page(text: str) -> BilingualText | None:
"""Deterministically parse the programme-page course/attendance fact tables.
The LLM extraction sees these tables only as fragmented visible text and
has produced lossy structure values (e.g. dropped the on-campus weeks), so
the parsed page block takes precedence. Returns None when the page does
not expose the attendance block, leaving the LLM value untouched.
"""
text = text or ''
campus = re.search(
r'class=["\']on-campus["\'][^>]*>\s*(\d+)\s*Wochen\s*<small>\s*Am\s+Campus',
text, flags=re.IGNORECASE | re.DOTALL,
)
if not campus:
return None
de_parts: list[str] = []
en_parts: list[str] = []
core = re.search(
r'class=["\']obligatory["\'].*?(\d+).*?Pflichtkurse',
text, flags=re.IGNORECASE | re.DOTALL,
)
if core:
de_parts.append(f"{core.group(1)} Pflichtkurse")
en_parts.append(f"{core.group(1)} core courses")
electives = re.search(
r'class=["\']optional["\'].*?(\d+).*?Wahlkurse',
text, flags=re.IGNORECASE | re.DOTALL,
)
if electives:
de_parts.append(f"{electives.group(1)} Wahlkurse")
en_parts.append(f"{electives.group(1)} electives")
de_parts.append(f"{campus.group(1)} Wochen am Campus")
en_parts.append(f"{campus.group(1)} weeks on campus")
abroad = re.search(
r'class=["\']outside-campus["\'][^>]*>\s*\+?\s*(\d+)\s*Wochen\s*<small>\s*im\s+Ausland',
text, flags=re.IGNORECASE | re.DOTALL,
)
if abroad:
de_parts.append(f"{abroad.group(1)} Wochen im Ausland")
en_parts.append(f"{abroad.group(1)} weeks abroad")
for extra in re.finditer(
r'class=["\']outside-campus["\'][^>]*>\s*\+?\s*<small>\s*([A-Za-zÄÖÜäöüß][A-Za-zÄÖÜäöüß -]*?)\s*</small>',
text, flags=re.IGNORECASE | re.DOTALL,
):
component_de = extra.group(1).strip()
de_parts.append(component_de)
en_parts.append(STRUCTURE_EXTRA_TRANSLATIONS.get(component_de, component_de))
return BilingualText(de=", ".join(de_parts), en=", ".join(en_parts))
def apply_deterministic_source_facts(extracted: AllProgrammesSchema, pages: dict[str, str]) -> AllProgrammesSchema:
"""Override LLM prose where the official page exposes a structured fact block."""
source_keys_by_programme = {
'emba': ['emba'],
'iemba': ['iemba', 'iemba_es'],
}
for programme_key, source_keys in source_keys_by_programme.items():
for source_key in source_keys:
locations = _extract_locations_from_programme_page(pages.get(source_key, ''))
if locations:
getattr(extracted, programme_key).locations = locations
break
for source_key in source_keys:
structure = _extract_structure_from_programme_page(pages.get(source_key, ''))
if structure:
getattr(extracted, programme_key).structure = structure
break
return extracted
def to_facts_document(extracted: AllProgrammesSchema) -> dict:
"""Convert the extraction schema into the programme_facts.json layout."""
def programme(p: ProgrammeFactsSchema, source_urls: list[str]) -> dict:
return {
'official_name': p.official_name,
'current_cohort': p.current_cohort,
'language': p.language.model_dump(),
'programme_start': p.programme_start,
'duration': p.duration.model_dump(),
'ects_credits': p.ects_credits,
'structure': p.structure.model_dump(),
'locations': p.locations.model_dump(),
'tuition_chf': {
'first_deadline': p.first_deadline.model_dump(),
'final_deadline': p.final_deadline.model_dump(),
'note': {
'de': 'Fristabhängige Studiengebühr: frühere Bewerbung = reduzierte Gebühr',
'en': 'Deadline-based tuition: earlier application = reduced fee',
},
},
'advisor': {
'name': p.advisor_name,
'email': p.advisor_email,
'phone': p.advisor_phone,
},
'source_urls': source_urls,
}
return {
'generated_at': date.today().isoformat(),
'generator': 'src/pipeline/update_programme_facts.py',
'sources': list(FACT_SOURCES.values()),
'programmes': {
'emba': programme(extracted.emba, [FACT_SOURCES['emba'], FACT_SOURCES['deadlines'], FACT_SOURCES['emba_plan']]),
'iemba': programme(extracted.iemba, [FACT_SOURCES['iemba'], FACT_SOURCES['iemba_es'], FACT_SOURCES['deadlines'], FACT_SOURCES['iemba_plan']]),
'emba_x': programme(extracted.emba_x, [FACT_SOURCES['emba_x'], FACT_SOURCES['deadlines']]),
},
}
# ----------------------------------- Diff ------------------------------------
DESCRIPTIVE_FACT_SUFFIXES = (
'duration.de',
'duration.en',
'structure.de',
'structure.en',
)
LOCATION_FACT_SUFFIXES = (
'locations.de',
'locations.en',
)
FACT_COMPARISON_STOP_WORDS = {
'a',
'am',
'and',
'as',
'at',
'auf',
'bis',
'by',
'das',
'der',
'die',
'en',
'for',
'im',
'in',
'max',
'maximum',
'mit',
'of',
'on',
'the',
'to',
'up',
'und',
'with',
}
FACT_SYNONYM_PHRASES = (
(r'\bpersonal\s+development\s+program(?:me)?\b', 'personal development'),
(r'\bpersonliche\s+entwicklung\b', 'personal development'),
(r'\bpersoenliche\s+entwicklung\b', 'personal development'),
(r'\bcapstone\s+projekt\b', 'capstone project'),
(r'\bselbststudium\b', 'self study'),
(r'\bself\s*study\b', 'self study'),
(r'\bpflichtkurse?n?\b', 'core courses'),
(r'\bwahlkurse?n?\b', 'electives'),
(r'\bessential\s+kurse?n?\b', 'essential courses'),
(r'\bwochen\s+am\s+campus\b', 'weeks on campus'),
(r'\bwochen\s+im\s+ausland\b', 'weeks abroad'),
(r'\bprogramm\b', 'program'),
(r'\bprogramme\b', 'program'),
)
STRUCTURE_COMPONENT_PATTERNS = {
'core_courses': r'\bcore\s+courses?\b',
'electives': r'\belectives?\b',
'campus_weeks': r'\bweeks?\s+on\s+campus\b',
'abroad_weeks': r'\bweeks?\s+abroad\b',
'capstone': r'\bcapstone\s+project\b',
'self_study': r'\bself\s+study\b',
'personal_development': r'\bpersonal\s+development\b',
'thesis': r'\b(?:thesis|diplomarbeit)\b',
'impact_projects': r'\bimpact\s+projects?\b',
'online': r'\bonline\b',
'essential_courses': r'\bessential\s+courses?\b',
}
def _flat_facts(d: dict, prefix: str = '') -> dict:
items = {}
for key, value in (d or {}).items():
flat_key = f"{prefix}{key}"
if isinstance(value, dict):
items.update(_flat_facts(value, flat_key + '.'))
elif not isinstance(value, list):
items[flat_key] = value
return items
def _set_nested_value(d: dict, dotted_key: str, value) -> None:
current = d
parts = dotted_key.split('.')
for part in parts[:-1]:
current = current[part]
current[parts[-1]] = value
def _strip_accents(value: str) -> str:
normalized = unicodedata.normalize('NFKD', value)
return ''.join(ch for ch in normalized if not unicodedata.combining(ch))
def _normalize_fact_phrases(value: str) -> str:
value = _strip_accents(str(value)).casefold()
value = value.replace('&', ' and ')
for pattern, replacement in FACT_SYNONYM_PHRASES:
value = re.sub(pattern, replacement, value, flags=re.IGNORECASE)
return value
def _canonical_text(value: str) -> str:
value = _normalize_fact_phrases(value)
value = re.sub(r'[^a-z0-9]+', ' ', value)
return re.sub(r'\s+', ' ', value).strip()
def _meaningful_tokens(value: str) -> set[str]:
return {
token
for token in _canonical_text(value).split()
if token not in FACT_COMPARISON_STOP_WORDS
}
def _number_signature(value: str) -> tuple[str, ...]:
return tuple(re.findall(r'\d+(?:\.\d+)?', str(value)))
def _structure_component_signature(value: str) -> set[str]:
normalized = _normalize_fact_phrases(value)
return {
component
for component, pattern in STRUCTURE_COMPONENT_PATTERNS.items()
if re.search(pattern, normalized, flags=re.IGNORECASE)
}
def _comparison_decision(
materially_changed: bool,
confidence: float,
reason: str,
fact_value,
preserve_existing: bool,
) -> FactComparisonDecision:
return FactComparisonDecision(
materially_changed=materially_changed,
confidence=confidence,
reason=reason,
fact_value='' if fact_value is None else str(fact_value),
preserve_existing=preserve_existing,
)
def _is_missing_extracted_value(fact_key: str, value) -> bool:
"""Return whether a schema value represents unavailable source data."""
if value is None:
return True
if isinstance(value, str):
return not value.strip()
if not isinstance(value, bool) and value == 0:
return fact_key.endswith(('.fee', 'ects_credits'))
return False
def _deterministic_fact_comparison(
fact_key: str,
existing_value,
observed_value,
) -> FactComparisonDecision | None:
if existing_value == observed_value:
return _comparison_decision(False, 1.0, "Values are identical.", existing_value, True)
existing_missing = _is_missing_extracted_value(fact_key, existing_value)
observed_missing = _is_missing_extracted_value(fact_key, observed_value)
if observed_missing:
return _comparison_decision(
False,
1.0,
"New extraction is missing; source absence is not evidence that the stored fact was removed.",
existing_value,
True,
)
if existing_missing:
return _comparison_decision(
True,
1.0,
"A previously missing fact is now available.",
observed_value,
False,
)
if not isinstance(existing_value, str) or not isinstance(observed_value, str):
return _comparison_decision(True, 1.0, "Structured or numeric value changed.", observed_value, False)
old_text = _canonical_text(existing_value)
new_text = _canonical_text(observed_value)
if old_text == new_text:
return _comparison_decision(
False,
1.0,
"Only punctuation, case, spelling, or separator formatting changed.",
existing_value,
True,
)
old_numbers = _number_signature(existing_value)
new_numbers = _number_signature(observed_value)
if old_numbers != new_numbers:
return _comparison_decision(True, 1.0, "Numeric/date signature changed.", observed_value, False)
if _is_location_fact(fact_key):
if _meaningful_tokens(existing_value) == _meaningful_tokens(observed_value):
return _comparison_decision(False, 1.0, "Location wording/order changed only.", existing_value, True)
return _comparison_decision(True, 0.95, "Location set changed.", observed_value, False)
if fact_key.endswith(('duration.de', 'duration.en')) and old_numbers:
return _comparison_decision(False, 0.95, "Duration wording changed but numbers are stable.", existing_value, True)
if fact_key.endswith(('structure.de', 'structure.en')):
old_components = _structure_component_signature(existing_value)
new_components = _structure_component_signature(observed_value)
if old_components != new_components:
return _comparison_decision(True, 0.95, "Programme structure component set changed.", observed_value, False)
old_tokens = _meaningful_tokens(existing_value)
new_tokens = _meaningful_tokens(observed_value)
if old_tokens == new_tokens or old_tokens.issubset(new_tokens):
return _comparison_decision(
False,
0.95,
"Structure wording changed without changing numbers or components.",
existing_value,
True,
)
return None
def _is_descriptive_fact(key: str) -> bool:
return key.endswith(DESCRIPTIVE_FACT_SUFFIXES)
def _is_location_fact(key: str) -> bool:
return key.endswith(LOCATION_FACT_SUFFIXES)
def _is_non_material_text_change(key: str, old_value, new_value) -> bool:
"""Detect LLM wording drift for descriptive fields.
The extraction is LLM-based, so prose fields can fluctuate between terse
and verbose wording. Alerts should be driven by stable core facts, not
punctuation, ordering, or added explanatory detail.
"""
if not isinstance(old_value, str) or not isinstance(new_value, str):
return False
old_text = _canonical_text(old_value)
new_text = _canonical_text(new_value)
if old_text == new_text:
return True
if _is_location_fact(key):
return _meaningful_tokens(old_value) == _meaningful_tokens(new_value)
if not _is_descriptive_fact(key):
return False
old_tokens = _meaningful_tokens(old_value)
new_tokens = _meaningful_tokens(new_value)
if old_tokens and old_tokens.issubset(new_tokens):
return True
if key.endswith(('duration.de', 'duration.en')):
old_numbers = _number_signature(old_value)
new_numbers = _number_signature(new_value)
return old_numbers == new_numbers and bool(old_numbers)
return False
def _is_material_change(key: str, old_value, new_value) -> bool:
if old_value == new_value:
return False
if _is_non_material_text_change(key, old_value, new_value):
return False
return True
def _source_keys_for_fact(programme_key: str, fact_key: str) -> list[str]:
if fact_key.startswith('tuition_chf.') or 'deadline' in fact_key:
return ['deadlines']
if programme_key == 'emba':
return ['emba', 'emba_plan']
if programme_key == 'iemba':
return ['iemba', 'iemba_es', 'iemba_plan']
if programme_key == 'emba_x':
return ['emba_x']
return list(FACT_SOURCES)
def _snippet_for_fact(pages: dict[str, str], source_keys: list[str], observed_value) -> str:
observed_tokens = [
token for token in _meaningful_tokens(str(observed_value))
if len(token) > 2
][:5]
snippets = []
for source_key in source_keys:
text = pages.get(source_key, '') or ''
if not text:
continue
canonical_text = _canonical_text(text)
if observed_tokens and not all(token in canonical_text for token in observed_tokens[:2]):
snippets.append(text[:3000])
continue
snippets.append(text[:3000])
return "\n\n".join(snippets)[:8000]
def evaluate_fact_against_existing(
existing_value,
page_content: str,
fact_key: str,
source_info: str,
language: str = '',
observed_value=None,
) -> FactComparisonDecision:
"""Decide whether an extracted value is a material change from storage."""
if observed_value is None:
observed_value = page_content
deterministic = _deterministic_fact_comparison(fact_key, existing_value, observed_value)
if deterministic is not None:
return deterministic
try:
from src.rag.models import ModelConfigurator
model = ModelConfigurator.get_main_agent_model().with_structured_output(
FactComparisonDecision
)
decision = model.invoke(FACT_COMPARISON_PROMPT.format(
fact_key=fact_key,
language=language or 'unknown',
source_info=source_info,
existing_value=existing_value,
observed_value=observed_value,
page_content=(page_content or '')[:8000],
))
if decision.materially_changed:
decision.preserve_existing = False
decision.fact_value = str(observed_value)
elif decision.preserve_existing:
decision.fact_value = str(existing_value)
return decision
except Exception as exc:
logger.warning(
"Could not run LLM fact comparison for %s; preserving existing value "
"to avoid an ambiguous overwrite: %s",
fact_key,
exc,
)
return _comparison_decision(
False,
0.0,
"LLM comparison unavailable; ambiguous change preserved existing value.",
existing_value,
True,
)
def preserve_materially_unchanged_extractions(
old: dict,
new: dict,
pages: dict[str, str] | None = None,
) -> dict:
"""Compare extracted facts against stored facts before final diffing."""
old_programmes = (old or {}).get('programmes', {})
pages = pages or {}
for prog_key, new_prog in new.get('programmes', {}).items():
old_prog = old_programmes.get(prog_key, {})
old_flat, new_flat = _flat_facts(old_prog), _flat_facts(new_prog)
for key in sorted(set(old_flat) & set(new_flat)):
if old_flat[key] == new_flat[key]:
continue
full_key = f"{prog_key}.{key}"
source_keys = _source_keys_for_fact(prog_key, key)
if not any((pages.get(source_key) or '').strip() for source_key in source_keys):
logger.info(
"Preserving existing %s: no usable source content was fetched.",
full_key,
)
_set_nested_value(new_prog, key, old_flat[key])
continue
source_info = ", ".join(FACT_SOURCES[source_key] for source_key in source_keys if source_key in FACT_SOURCES)
decision = evaluate_fact_against_existing(
existing_value=old_flat[key],
observed_value=new_flat[key],
page_content=_snippet_for_fact(pages, source_keys, new_flat[key]),
fact_key=full_key,
source_info=source_info,
language='de' if key.endswith('.de') else 'en' if key.endswith('.en') else '',
)
if decision.preserve_existing or not decision.materially_changed:
logger.info(
"Preserving existing %s: %s",
full_key,
decision.reason,
)
_set_nested_value(new_prog, key, old_flat[key])
return new
def preserve_non_material_changes(old: dict, new: dict) -> dict:
"""Keep existing wording when the new extraction is only a paraphrase."""
old_programmes = (old or {}).get('programmes', {})
for prog_key, new_prog in new.get('programmes', {}).items():
old_prog = old_programmes.get(prog_key, {})
old_flat, new_flat = _flat_facts(old_prog), _flat_facts(new_prog)
for key in sorted(set(old_flat) & set(new_flat)):
if old_flat[key] == new_flat[key]:
continue
full_key = f"{prog_key}.{key}"
if not _is_material_change(full_key, old_flat[key], new_flat[key]):
_set_nested_value(new_prog, key, old_flat[key])
return new
def diff_facts(old: dict, new: dict) -> list[str]:
"""Compare volatile values between old and new facts; returns change lines."""
changes = []
old_programmes = (old or {}).get('programmes', {})
for prog_key, new_prog in new.get('programmes', {}).items():
old_prog = old_programmes.get(prog_key, {})
old_flat, new_flat = _flat_facts(old_prog), _flat_facts(new_prog)
for key in sorted(set(old_flat) | set(new_flat)):
full_key = f"{prog_key}.{key}"
if _is_material_change(full_key, old_flat.get(key), new_flat.get(key)):
changes.append(
f"{prog_key}.{key}: {old_flat.get(key, '<missing>')} -> {new_flat.get(key, '<missing>')}"
)
return changes
def notify_changes(changes: list[str]) -> None:
try:
from src.notification.notification_center import NotificationCenter
NotificationCenter().send_notification(
subject="Programme facts changed on official websites",
body="The fact checker detected changes:\n\n" + "\n".join(changes),
channel="all",
)
except Exception as e:
logger.warning(f"Could not send change notification: {e}")
# ----------------------------------- Main ------------------------------------
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--dry-run', action='store_true', help='Show diff without writing')
args = parser.parse_args()
old_facts = {}
if os.path.exists(FACTS_PATH):
with open(FACTS_PATH, encoding='utf-8') as f:
old_facts = json.load(f)
pages = fetch_sources()
try:
extracted = extract_facts(pages, existing_facts=old_facts)
except Exception as exc:
logger.error(f"Could not extract programme facts; existing facts file was not changed: {exc}")
return 1
extracted = apply_deterministic_fallbacks(extracted, pages)
extracted = apply_deterministic_source_facts(extracted, pages)
new_facts = to_facts_document(extracted)
if old_facts:
new_facts = preserve_materially_unchanged_extractions(old_facts, new_facts, pages)
new_facts = preserve_non_material_changes(old_facts, new_facts)
changes = diff_facts(old_facts, new_facts)
if changes:
logger.warning(f"Detected {len(changes)} fact change(s):")
for change in changes:
logger.warning(f" {change}")
else:
logger.info("No fact changes detected.")
if args.dry_run:
print(json.dumps(new_facts, indent=2, ensure_ascii=False))
return 0
if old_facts and not changes:
logger.info("Keeping existing facts file because only non-material wording changed.")
return 0
os.makedirs(os.path.dirname(FACTS_PATH), exist_ok=True)
with open(FACTS_PATH, 'w', encoding='utf-8') as f:
json.dump(new_facts, f, indent=2, ensure_ascii=False)
logger.info(f"Wrote {FACTS_PATH}")
if changes:
notify_changes(changes)
# Invalidate the in-process cache so a running app picks up new facts
try:
from src.rag.verified_facts import VerifiedFacts
VerifiedFacts.reset_cache()
except Exception:
pass
return 0
if __name__ == '__main__':
sys.exit(main())
|