Spaces:
Sleeping
Sleeping
File size: 63,333 Bytes
6b6d1fd | 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 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 | # -*- coding: utf-8 -*-
"""Scimplify.ipynb
Automatically generated by Colab.
Original file is located at
https://colab.research.google.com/drive/11L85VXrmvxrfXd6A9FGuJjI53nVtM0tN
# Scimplify
A NeuroAI paper simplifier. You paste a paragraph and get a plain-language
explanation back, with citations to the retrieved chunks the explanation
came from. The system refuses to answer if it can't ground the claims.
## 1. Setup
"""
import os, json, re, io, textwrap, time
from pathlib import Path
from collections import Counter, defaultdict
from typing import List, Dict, Tuple, Optional
from concurrent.futures import ThreadPoolExecutor, as_completed
import requests
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import openai
import chromadb
from chromadb.utils import embedding_functions
import gradio as gr
from sentence_transformers import SentenceTransformer
from PyPDF2 import PdfReader
try:
import os
api_key = os.getenv("OPENAI_API_KEY")
except (ImportError, Exception):
assert os.environ.get("OPENAI_API_KEY"), "Set OPENAI_API_KEY env var"
client_oai = openai.OpenAI()
GENERATOR_MODEL = "gpt-4o-mini"
JUDGE_MODEL = "gpt-4o-mini"
GENERATOR_TEMPERATURE = 0.2
JUDGE_TEMPERATURE = 0.3
JUDGE_N_SAMPLES = 3
BOOTSTRAP_N = 2000
BOOTSTRAP_ALPHA = 0.05
_rng = np.random.default_rng(7)
RUN_EXPERIMENTS = False # re-run experiments
LIVE_SEMANTIC_CHECK = True # adds 1s per query
JUDGE_PARALLELISM = 2 # rate limit cap
print(f"generator: {GENERATOR_MODEL}")
print(f"judge: {JUDGE_MODEL}")
print(f"experiments: {'WILL RE-RUN' if RUN_EXPERIMENTS else 'using cached results'}")
print(f"live semantic check: {'on' if LIVE_SEMANTIC_CHECK else 'off'}")
"""## 2. Data loading"""
REPO_RAW_BASE = "https://raw.githubusercontent.com/martazavro/scimplify_data/main"
LOCAL_DATA_DIR = Path("./data")
def _load_json(filename):
url = f"{REPO_RAW_BASE}/{filename}"
try:
r = requests.get(url, timeout=10)
r.raise_for_status()
print(f"loaded {filename} from repo")
return r.json()
except Exception as e:
local = LOCAL_DATA_DIR / filename
if local.exists():
print(f"repo fetch failed ({e.__class__.__name__}), loaded {filename} from local")
return json.loads(local.read_text())
raise FileNotFoundError(
f"Could not load {filename}. Set REPO_RAW_BASE correctly "
f"or place the file in ./data/{filename}"
)
neuroai_concepts = _load_json("concepts.json")
print(f"loaded {len(neuroai_concepts)} concepts")
def validate_validation_set(vs):
items = vs["items"]
ids = [x["id"] for x in items]
assert len(set(ids)) == len(ids), "duplicate ids"
required = {"id", "passage", "source", "key_terms", "category", "difficulty", "reference_explanation"}
valid_cats = {"concepts_only", "recent_paper", "both", "neither"}
for item in items:
missing = required - set(item.keys())
assert not missing, f"item {item.get('id')} missing {missing}"
assert item["category"] in valid_cats
cat_counts = Counter(x["category"] for x in items)
print(f"validation set: {len(items)} items")
print(f" by category: {dict(cat_counts)}")
validation_set = _load_json("validation_set.json")
validate_validation_set(validation_set)
"""## 3. PDF extraction and chunking"""
def extract_text_from_pdf(pdf_file):
reader = PdfReader(pdf_file)
text = ""
for page in reader.pages:
page_text = page.extract_text()
if page_text:
text += page_text + "\n"
return text.strip()
def chunk_text(text, chunk_size=300, overlap=50):
paragraphs = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
chunks, current, current_len = [], [], 0
for para in paragraphs:
words = para.split()
n = len(words)
if n > chunk_size:
if current:
chunks.append(" ".join(current))
tail = current[-overlap:] if len(current) > overlap else current
current = list(tail); current_len = len(current)
for i in range(0, n, chunk_size - overlap):
chunk = words[i:i+chunk_size]
if len(chunk) > 30:
chunks.append(" ".join(chunk))
current = []; current_len = 0
elif current_len + n > chunk_size:
chunks.append(" ".join(current))
tail = current[-overlap:] if len(current) > overlap else current
current = list(tail) + words; current_len = len(current)
else:
current.extend(words); current_len += n
if current and len(current) > 30:
chunks.append(" ".join(current))
return chunks
"""## 4. Vector store setup"""
chroma_client = chromadb.Client()
ef = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="all-MiniLM-L6-v2")
def reset_concepts_collection():
try:
chroma_client.delete_collection("neuroai_concepts")
except Exception:
pass
coll = chroma_client.create_collection(name="neuroai_concepts", embedding_function=ef)
for entry in neuroai_concepts:
doc = (
f"Concept: {entry['concept']}\n"
f"Definition: {entry['definition']}\n"
f"Context: {entry['context']}\n"
f"Typically found in: {entry['typical_usage']}"
)
coll.add(
documents=[doc],
ids=[entry["id"]],
metadatas=[{"concept_name": entry["concept"], "concept_id": entry["id"]}]
)
return coll
def reset_papers_collection():
try:
chroma_client.delete_collection("neuroai_papers")
except Exception:
pass
return chroma_client.create_collection(name="neuroai_papers", embedding_function=ef)
concepts_collection = reset_concepts_collection()
papers_collection = reset_papers_collection()
print(f"concepts: {concepts_collection.count()}, papers: {papers_collection.count()}")
"""## 5. Recent papers ingestion"""
PAPER_CHUNKS_URL = f"{REPO_RAW_BASE}/paper_chunks.json"
def load_paper_chunks():
r = requests.get(PAPER_CHUNKS_URL, timeout=15)
r.raise_for_status()
return r.json()
def ingest_paper_chunks_from_json():
chunks = load_paper_chunks()
if not chunks:
print("paper_chunks.json was empty")
return 0
documents = [c["text"] for c in chunks]
ids = [c["chunk_id"] for c in chunks]
metadatas = [{
"source_name": c["source_name"],
"source_type": c["source_type"],
"arxiv_id": c["arxiv_id"],
"title": c["title"],
"chunk_idx": c["chunk_idx"],
"chunk_id": c["chunk_id"],
} for c in chunks]
papers_collection.add(documents=documents, ids=ids, metadatas=metadatas)
by_paper = {}
for c in chunks:
by_paper[c["arxiv_id"]] = by_paper.get(c["arxiv_id"], 0) + 1
for aid, n in by_paper.items():
print(f" {aid}: {n} chunks")
print(f"papers_collection now has {papers_collection.count()} total chunks")
return len(chunks)
ingest_paper_chunks_from_json()
"""## 6. arXiv ingestion"""
import arxiv
def _existing_arxiv_ids():
if papers_collection.count() == 0:
return set()
metas = papers_collection.get()["metadatas"]
return {m.get("arxiv_id") for m in metas if m.get("arxiv_id")}
def ingest_from_arxiv(query="neuroAI OR (neural AND brain AND deep learning)",
max_results=10,
sort_by_recent=True,
verbose=True):
"""Search arXiv, download PDFs, chunk them, add to papers_collection.
Returns dict with stats: {n_papers, n_chunks, n_skipped, errors}.
Already-ingested papers (matched by arxiv_id) are skipped.
"""
sort_by = arxiv.SortCriterion.SubmittedDate if sort_by_recent else arxiv.SortCriterion.Relevance
arxiv_client = arxiv.Client(page_size=20, delay_seconds=3.0, num_retries=3)
search = arxiv.Search(query=query, max_results=max_results, sort_by=sort_by)
existing = _existing_arxiv_ids()
download_dir = Path("./arxiv_papers")
download_dir.mkdir(exist_ok=True)
n_papers, n_chunks, n_skipped = 0, 0, 0
errors = []
for result in arxiv_client.results(search):
# arxiv.org/abs/2509.23566v1 -> "2509.23566"
full_id = result.entry_id.rsplit("/", 1)[-1]
arxiv_id = full_id.split("v")[0]
if arxiv_id in existing:
n_skipped += 1
if verbose:
print(f" skip {arxiv_id} (already ingested)")
continue
try:
if verbose:
print(f" fetching {arxiv_id}: {result.title[:60]}...")
pdf_path = result.download_pdf(dirpath=str(download_dir),
filename=f"{arxiv_id}.pdf")
text = extract_text_from_pdf(pdf_path)
chunks = chunk_text(text)
if not chunks:
errors.append(f"{arxiv_id}: no chunks extracted")
continue
chunk_ids = [f"arxiv_{arxiv_id.replace('.', '_')}::c{i}" for i in range(len(chunks))]
metadatas = [{
"source_name": result.title,
"source_type": "arxiv_paper",
"arxiv_id": arxiv_id,
"title": result.title,
"chunk_idx": i,
"chunk_id": chunk_ids[i],
} for i in range(len(chunks))]
papers_collection.add(documents=chunks, ids=chunk_ids, metadatas=metadatas)
existing.add(arxiv_id) # avoid double-add within the same call
n_papers += 1
n_chunks += len(chunks)
if verbose:
print(f" -> added {len(chunks)} chunks")
except Exception as e:
errors.append(f"{arxiv_id}: {e.__class__.__name__}: {e}")
if verbose:
print(f" ERROR: {e}")
summary = {
"n_papers": n_papers,
"n_chunks": n_chunks,
"n_skipped": n_skipped,
"errors": errors,
"total_in_kb": papers_collection.count(),
}
if verbose:
print(f"\ningested {n_papers} papers ({n_chunks} chunks), skipped {n_skipped} duplicates")
if errors:
print(f"errors: {len(errors)}")
print(f"total in knowledge base: {summary['total_in_kb']} chunks")
return summary
ingest_from_arxiv(query="NeuroAI", max_results=2)
ingest_from_arxiv(query="NeuroAI", max_results=15)
"""## 7. Retrieval variants"""
def _flexible_last_word(word):
if len(word) < 4:
return re.escape(word)
stem = re.escape(word[:-2])
return stem + r"[a-z]{0,4}"
def build_concept_patterns(concept_entry):
name = concept_entry["concept"]
abbrev_match = re.search(r"\(([^)]+)\)", name)
abbrev = abbrev_match.group(1) if abbrev_match else None
base = re.sub(r"\s*\([^)]+\)", "", name).strip()
patterns = []
words = base.split()
if len(words) == 1:
long_re = r"\b" + _flexible_last_word(words[0]) + r"\b"
else:
parts = [re.escape(w) for w in words[:-1]] + [_flexible_last_word(words[-1])]
long_re = r"\b" + r"\s+".join(parts) + r"\b"
patterns.append(re.compile(long_re, re.IGNORECASE))
if abbrev:
patterns.append(re.compile(r"\b" + re.escape(abbrev) + r"s?\b"))
return patterns
CONCEPT_PATTERNS = [(entry, build_concept_patterns(entry)) for entry in neuroai_concepts]
def _concept_doc_text(entry):
return (
f"Concept: {entry['concept']}\n"
f"Definition: {entry['definition']}\n"
f"Context: {entry['context']}\n"
f"Typically found in: {entry['typical_usage']}"
)
def regex_retrieve(passage):
hits = []
for entry, patterns in CONCEPT_PATTERNS:
if any(p.search(passage) for p in patterns):
hits.append({
"type": "regex_concept",
"concept_name": entry["concept"],
"concept_id": entry["id"],
"chunk_id": entry["id"],
"content": _concept_doc_text(entry),
"distance": 0.0,
"source_method": "regex",
})
return hits
def retrieve_concepts_embedding(passage, n_results=3):
results = concepts_collection.query(query_texts=[passage], n_results=n_results)
out = []
for doc, meta, dist in zip(results["documents"][0], results["metadatas"][0], results["distances"][0]):
out.append({
"type": "concept",
"concept_name": meta["concept_name"],
"concept_id": meta.get("concept_id"),
"chunk_id": meta.get("concept_id"),
"content": doc,
"distance": round(dist, 3),
"source_method": "embedding",
})
return out
def retrieve_paper_chunks(passage, n_results=3):
if papers_collection.count() == 0:
return []
results = papers_collection.query(
query_texts=[passage],
n_results=min(n_results, papers_collection.count())
)
out = []
for doc, meta, dist in zip(results["documents"][0], results["metadatas"][0], results["distances"][0]):
out.append({
"type": "paper_chunk",
"source_name": meta["source_name"],
"source_type": meta["source_type"],
"chunk_id": meta.get("chunk_id"),
"content": doc,
"distance": round(dist, 3),
"source_method": "embedding",
})
return out
def hybrid_retrieve_concepts(passage, n_embedding=3, max_total=6):
rgx = regex_retrieve(passage)
seen_names = {h["concept_name"] for h in rgx}
out = list(rgx)
if len(out) < max_total:
emb = retrieve_concepts_embedding(passage, n_results=n_embedding)
for hit in emb:
if hit["concept_name"] not in seen_names:
out.append(hit)
seen_names.add(hit["concept_name"])
else:
for r in out:
if r["concept_name"] == hit["concept_name"]:
r["source_method"] = "both"
break
if len(out) >= max_total:
break
return out
def retrieve_for_variant(passage, variant, n_concepts=3, n_papers=3):
if variant == "no_rag":
return [], []
elif variant == "embedding_only":
return (retrieve_concepts_embedding(passage, n_results=n_concepts),
retrieve_paper_chunks(passage, n_results=n_papers))
elif variant == "regex_only":
return (regex_retrieve(passage),
retrieve_paper_chunks(passage, n_results=n_papers))
elif variant == "hybrid":
return (hybrid_retrieve_concepts(passage, n_embedding=n_concepts),
retrieve_paper_chunks(passage, n_results=n_papers))
else:
raise ValueError(f"unknown variant: {variant}")
"""## 8. Citation-enforced generation with semantic guard
"""
CITED_SYSTEM_PROMPT = """You are a scientific reading assistant that helps people understand passages from NeuroAI research papers.
You have access to retrieved context. Each source has a stable ID in square brackets like [c004] (for a concept definition) or [arxiv_2511_12345::c3] (for a paper chunk).
Your job:
1. Read the passage.
2. Rewrite it in plain language an undergraduate could follow.
3. For EVERY factual sentence in your explanation, append one or more citations in square brackets, drawn ONLY from the IDs of the retrieved sources shown to you.
4. Do not invent citation IDs. Do not cite sources you were not shown.
5. If the retrieved context does not contain enough information to answer faithfully, output EXACTLY this string and nothing else:
I don't have enough evidence in the retrieved context.
Format:
**Key terms:** short definitions of technical terms, each with its citation
**Plain-language version:** the passage rewritten clearly, with citations on every factual sentence
**What this means in context:** 1-2 sentences on why this matters, with citations
"""
ABSTAIN_MESSAGE = "I don't have enough evidence in the retrieved context."
CITATION_PATTERN = re.compile(r"\[([a-zA-Z0-9_\-:]+)\]")
SEMANTIC_FAIL_THRESHOLD = 0.5
def _format_context_block(concept_results, paper_results):
lines = []
if concept_results:
lines.append("CONCEPT DEFINITIONS:")
for r in concept_results:
cid = r.get("chunk_id") or r.get("concept_id")
lines.append(f"\n[{cid}] {r['content']}")
lines.append("---")
if paper_results:
lines.append("\nPAPER/ARTICLE CONTEXT:")
for r in paper_results:
cid = r.get("chunk_id")
lines.append(f"\n[{cid}] (from {r['source_name']}): {r['content']}")
lines.append("---")
if not concept_results and not paper_results:
lines.append("(no context retrieved)")
return "\n".join(lines)
def _collect_allowed_ids(concept_results, paper_results):
ids = set()
for r in concept_results + paper_results:
cid = r.get("chunk_id") or r.get("concept_id")
if cid:
ids.add(cid)
return ids
def _build_chunk_lookup(concept_results, paper_results):
"""Map citation_id -> chunk content. Used by the semantic check."""
lookup = {}
for r in concept_results + paper_results:
cid = r.get("chunk_id") or r.get("concept_id")
if cid:
lookup[cid] = r["content"]
return lookup
def generate_cited_explanation(passage, concept_results, paper_results, model=None):
model = model or GENERATOR_MODEL
context_block = _format_context_block(concept_results, paper_results)
user_msg = f"{context_block}\n\nPASSAGE TO EXPLAIN:\n{passage}"
resp = client_oai.chat.completions.create(
model=model,
temperature=GENERATOR_TEMPERATURE,
messages=[
{"role": "system", "content": CITED_SYSTEM_PROMPT},
{"role": "user", "content": user_msg},
],
)
return resp.choices[0].message.content
def validate_citations(answer: str, allowed_ids: set) -> Tuple[bool, List[str]]:
"""Lexical guard: every citation ID in the answer must be in the allowed set."""
if ABSTAIN_MESSAGE in answer:
return True, []
cited = CITATION_PATTERN.findall(answer)
issues = []
if not cited:
issues.append("No citations found in non-abstain answer")
for cid in cited:
if cid not in allowed_ids:
issues.append(f"Invalid citation: {cid}")
return len(issues) == 0, issues
def _split_into_sentences(text):
"""Cheap sentence splitter that keeps citation brackets attached."""
# split on . ! ? followed by space and a capital, keeping the punctuation
parts = re.split(r"(?<=[.!?])\s+(?=[A-Z*])", text.strip())
return [p.strip() for p in parts if p.strip()]
def _strip_citations(sentence):
return CITATION_PATTERN.sub("", sentence).strip()
def check_sentence_supported(sentence_text, cited_chunks):
claim = _strip_citations(sentence_text)
if len(claim) < 10 or not cited_chunks:
return {"label": "skipped", "reason": "no claim or no chunks"}
evidence = "\n\n".join(f"[{cid}]: {text}" for cid, text in cited_chunks)
return verify_claim_against_evidence(claim, [evidence])
def semantic_per_sentence_check(answer, chunk_lookup):
if ABSTAIN_MESSAGE in answer:
return []
sentences = _split_into_sentences(answer)
findings = []
for sent in sentences:
cited_ids = CITATION_PATTERN.findall(sent)
if not cited_ids:
continue
cited_chunks = [(cid, chunk_lookup[cid]) for cid in cited_ids if cid in chunk_lookup]
if not cited_chunks:
continue
result = check_sentence_supported(sent, cited_chunks)
findings.append({
"sentence": sent,
"citations": cited_ids,
"label": result["label"],
"reason": result["reason"],
})
return findings
def annotate_unsupported_sentences(answer, findings):
"""Mark unsupported sentences in the rendered output."""
for f in findings:
if f["label"] in ("contradicted", "insufficient"):
marker = "⚠️ "
if marker not in f["sentence"]:
answer = answer.replace(f["sentence"], marker + f["sentence"], 1)
return answer
def generate_with_citation_guard(passage, concept_results, paper_results, model=None,
allow_no_context_bypass=False,
do_semantic_check=None):
do_semantic_check = (do_semantic_check if do_semantic_check is not None
else LIVE_SEMANTIC_CHECK)
if allow_no_context_bypass and not concept_results and not paper_results:
resp = client_oai.chat.completions.create(
model=model or GENERATOR_MODEL,
temperature=GENERATOR_TEMPERATURE,
messages=[
{"role": "system", "content": "You are a scientific reading assistant. Explain the given passage in plain language that an undergraduate could follow. Be concise."},
{"role": "user", "content": f"PASSAGE:\n{passage}"},
],
)
return {
"answer": resp.choices[0].message.content,
"valid_citations": None,
"guard_triggered": False,
"issues": [],
"abstained": False,
"semantic_findings": [],
"semantic_fail_rate": np.nan,
}
raw = generate_cited_explanation(passage, concept_results, paper_results, model=model)
allowed_ids = _collect_allowed_ids(concept_results, paper_results)
ok, issues = validate_citations(raw, allowed_ids)
# lexical guard
if not ok:
return {
"answer": ABSTAIN_MESSAGE,
"valid_citations": False,
"guard_triggered": True,
"issues": issues,
"abstained": True,
"raw_rejected": raw,
"semantic_findings": [],
"semantic_fail_rate": np.nan,
}
# semantic per-sentence check
findings = []
semantic_fail_rate = np.nan
if do_semantic_check and ABSTAIN_MESSAGE not in raw:
chunk_lookup = _build_chunk_lookup(concept_results, paper_results)
findings = semantic_per_sentence_check(raw, chunk_lookup)
if findings:
n_failed = sum(1 for f in findings if f["label"] in ("contradicted", "insufficient"))
semantic_fail_rate = n_failed / len(findings)
if semantic_fail_rate > SEMANTIC_FAIL_THRESHOLD:
return {
"answer": ABSTAIN_MESSAGE,
"valid_citations": True,
"guard_triggered": True,
"issues": [f"semantic check failed: {n_failed}/{len(findings)} sentences unsupported"],
"abstained": True,
"raw_rejected": raw,
"semantic_findings": findings,
"semantic_fail_rate": semantic_fail_rate,
}
raw = annotate_unsupported_sentences(raw, findings)
return {
"answer": raw,
"valid_citations": True,
"guard_triggered": False,
"issues": [],
"abstained": ABSTAIN_MESSAGE in raw,
"semantic_findings": findings,
"semantic_fail_rate": semantic_fail_rate,
}
"""## 9. LLM-as-judge metrics
"""
def _coerce_score(x):
try:
v = int(float(x))
except Exception:
v = 0
return max(0, min(2, v))
def _single_judge_call(system_prompt, user_prompt):
try:
resp = client_oai.chat.completions.create(
model=JUDGE_MODEL,
temperature=JUDGE_TEMPERATURE,
response_format={"type": "json_object"},
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
],
)
data = json.loads(resp.choices[0].message.content)
return {
"score": _coerce_score(data.get("score", 0)),
"reason": str(data.get("reason", "")).strip(),
}
except Exception as e:
return {"score": None, "reason": f"ERROR: {e}"}
def _judge_call_parallel(system_prompt, user_prompt, n=None):
"""Run n judge calls in parallel via ThreadPoolExecutor."""
n = n or JUDGE_N_SAMPLES
results = [None] * n
with ThreadPoolExecutor(max_workers=min(n, JUDGE_PARALLELISM)) as ex:
futures = {ex.submit(_single_judge_call, system_prompt, user_prompt): i
for i in range(n)}
for fut in as_completed(futures):
i = futures[fut]
results[i] = fut.result()
return results
def _aggregate(runs):
valid = [r for r in runs if r["score"] is not None]
if not valid:
return {"score": None, "reasons": [r["reason"] for r in runs], "n_valid": 0}
return {
"score": sum(r["score"] for r in valid) / len(valid),
"reasons": [r["reason"] for r in valid],
"n_valid": len(valid),
}
CORRECTNESS_SYSTEM = """You are evaluating answer correctness for a question about a NeuroAI paper passage.
Given a passage, a reference explanation (gold-standard), and a system explanation, score the system explanation's correctness using ONLY the information in the passage and reference.
Return ONLY a JSON object:
{"score": <int 0/1/2>, "reason": "<one sentence>"}
Scoring scale:
- 0 = wrong (contradicts the passage or says something incorrect)
- 1 = partly correct (captures some but not all of the main idea, or adds unsupported claims)
- 2 = correct (faithful to what the passage actually says)
"""
def score_correctness(passage, reference, candidate):
user = f"PASSAGE:\n{passage}\n\nREFERENCE:\n{reference}\n\nSYSTEM EXPLANATION:\n{candidate}"
runs = _judge_call_parallel(CORRECTNESS_SYSTEM, user)
return _aggregate(runs)
EVIDENCE_SYSTEM = """You are evaluating whether a system explanation's key claims are supported by retrieved context.
Given a passage, the retrieved context that was shown to the system, and the system's explanation, score whether the explanation's factual claims are well-supported by the retrieved context.
Return ONLY a JSON object:
{"score": <int 0/1/2>, "reason": "<one sentence>"}
Scoring scale:
- 0 = unsupported (most claims cannot be found in retrieved context)
- 1 = partly supported (some claims supported, others require outside knowledge)
- 2 = well supported (claims are traceable to retrieved context)
If the retrieved context is empty (no RAG baseline), score 0.
"""
def score_evidence_support(passage, retrieved_context, candidate):
user = f"PASSAGE:\n{passage}\n\nRETRIEVED CONTEXT:\n{retrieved_context}\n\nSYSTEM EXPLANATION:\n{candidate}"
runs = _judge_call_parallel(EVIDENCE_SYSTEM, user)
return _aggregate(runs)
CITATION_SYSTEM = """You are evaluating whether citations in a system explanation are faithful.
The system was asked to cite each factual sentence with an ID from the retrieved context (like [c004] or [arxiv_2511_12345::c3]). Given the retrieved context and the system explanation with citations, score whether the citations are relevant and the cited material actually supports the adjacent claim.
Return ONLY a JSON object:
{"score": <int 0/1/2>, "reason": "<one sentence>"}
Scoring scale:
- 0 = unfaithful (citations invented, missing, or do not support adjacent claims)
- 1 = mixed (some citations support their claims, others do not)
- 2 = faithful (citations are present, relevant, and support adjacent claims)
If the answer is the abstention message ("I don't have enough evidence..."), score 2 (correctly declined).
"""
def score_citation_faithfulness(retrieved_context, candidate):
user = f"RETRIEVED CONTEXT:\n{retrieved_context}\n\nSYSTEM EXPLANATION:\n{candidate}"
runs = _judge_call_parallel(CITATION_SYSTEM, user)
return _aggregate(runs)
def score_all_metrics(passage, reference, retrieved_context, candidate):
"""Run all three metrics in parallel."""
with ThreadPoolExecutor(max_workers=3) as ex:
f_c = ex.submit(score_correctness, passage, reference, candidate)
f_e = ex.submit(score_evidence_support, passage, retrieved_context, candidate)
f_f = ex.submit(score_citation_faithfulness, retrieved_context, candidate)
return {
"correctness": f_c.result(),
"evidence_support": f_e.result(),
"citation_faithfulness": f_f.result(),
}
"""## 10. Claim-based faithfulness
"""
CLAIM_EXTRACTION_SYSTEM = """Extract atomic factual claims from the given answer.
Return ONLY a JSON object:
{"claims": ["claim 1", "claim 2", ...]}
Rules:
- Each claim should be a single, minimal factual assertion
- Ignore pure formatting, headers, or meta-commentary
- Skip citation markers like [c004] when extracting claims
- If there are no factual claims, return {"claims": []}
"""
EVIDENCE_EXTRACTION_SYSTEM = """Extract factual assertions from the given text chunk.
Return ONLY a JSON object:
{"assertions": ["assertion 1", "assertion 2", ...]}
Rules:
- One atomic factual assertion per entry
- Skip anything that is a question, opinion, or example
- If there are no assertions, return {"assertions": []}
"""
CLAIM_VERIFICATION_SYSTEM = """Classify if a claim is supported, contradicted, or insufficient given evidence.
Return ONLY a JSON object:
{"label": "supported" | "contradicted" | "insufficient", "reason": "<one short sentence>"}
Definitions:
- supported: the evidence directly supports the claim
- contradicted: the evidence contradicts the claim
- insufficient: the evidence is silent or unclear on the claim
"""
def _json_call(system_prompt, user_prompt, model=None):
model = model or JUDGE_MODEL
resp = client_oai.chat.completions.create(
model=model,
temperature=JUDGE_TEMPERATURE,
response_format={"type": "json_object"},
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
],
)
try:
return json.loads(resp.choices[0].message.content)
except Exception:
return {}
def extract_claims(answer):
data = _json_call(CLAIM_EXTRACTION_SYSTEM, f"ANSWER:\n{answer}")
return [c for c in data.get("claims", []) if c and isinstance(c, str)]
_ASSERTION_CACHE = {}
def extract_assertions_from_chunk(chunk):
key = hash(chunk)
if key in _ASSERTION_CACHE:
return _ASSERTION_CACHE[key]
data = _json_call(EVIDENCE_EXTRACTION_SYSTEM, f"CHUNK:\n{chunk}")
out = [a for a in data.get("assertions", []) if a and isinstance(a, str)]
_ASSERTION_CACHE[key] = out
return out
def _normalize_label(label):
x = (label or "").strip().lower()
if "support" in x: return "supported"
if "contrad" in x: return "contradicted"
return "insufficient"
def verify_claim_against_evidence(claim, assertions):
evidence_blob = "\n".join(assertions) if assertions else "NO_EVIDENCE"
data = _json_call(
CLAIM_VERIFICATION_SYSTEM,
f"CLAIM:\n{claim}\n\nEVIDENCE:\n{evidence_blob}"
)
return {
"label": _normalize_label(data.get("label")),
"reason": str(data.get("reason", "")).strip(),
}
def claim_based_faithfulness(answer, retrieved_chunks):
if ABSTAIN_MESSAGE in answer:
return {
"n_claims": 0,
"support_rate": np.nan,
"contradiction_rate": np.nan,
"unsupported_rate": np.nan,
"abstained": True,
"details": [],
}
claims = extract_claims(answer)
if not claims:
return {
"n_claims": 0,
"support_rate": np.nan,
"contradiction_rate": np.nan,
"unsupported_rate": np.nan,
"abstained": False,
"details": [],
}
with ThreadPoolExecutor(max_workers=JUDGE_PARALLELISM) as ex:
all_assertions_lists = list(ex.map(extract_assertions_from_chunk, retrieved_chunks))
all_assertions = [a for sub in all_assertions_lists for a in sub]
with ThreadPoolExecutor(max_workers=JUDGE_PARALLELISM) as ex:
verify_results = list(ex.map(
lambda c: verify_claim_against_evidence(c, all_assertions),
claims
))
labels = [r["label"] for r in verify_results]
details = [{"claim": c, **r} for c, r in zip(claims, verify_results)]
n = len(labels)
return {
"n_claims": n,
"support_rate": sum(1 for l in labels if l == "supported") / n,
"contradiction_rate": sum(1 for l in labels if l == "contradicted") / n,
"unsupported_rate": sum(1 for l in labels if l == "insufficient") / n,
"abstained": False,
"details": details,
}
"""## 11. Retrieval precision@k / recall@k and bootstrap CIs"""
def precision_recall_at_k(retrieved_chunks, gold_facts, k=3):
if not gold_facts:
return np.nan, np.nan
top_k = retrieved_chunks[:k]
if not top_k:
return 0.0, 0.0
rel_flags = []
for chunk in top_k:
c = chunk.lower()
is_rel = any(fact.lower() in c for fact in gold_facts)
rel_flags.append(is_rel)
precision = float(np.mean(rel_flags))
covered = 0
for fact in gold_facts:
if any(fact.lower() in chunk.lower() for chunk in top_k):
covered += 1
recall = covered / len(gold_facts)
return precision, recall
def bootstrap_ci(values, n_boot=None, alpha=None):
n_boot = n_boot or BOOTSTRAP_N
alpha = alpha or BOOTSTRAP_ALPHA
values = np.array(values, dtype=float)
values = values[~np.isnan(values)]
if len(values) == 0:
return np.nan, np.nan, np.nan
boots = np.empty(n_boot)
n = len(values)
for i in range(n_boot):
sample = _rng.choice(values, size=n, replace=True)
boots[i] = sample.mean()
lo = np.percentile(boots, 100 * (alpha / 2))
hi = np.percentile(boots, 100 * (1 - alpha / 2))
return float(values.mean()), float(lo), float(hi)
def format_ci(values, digits=3):
m, lo, hi = bootstrap_ci(values)
return f"{m:.{digits}f} [{lo:.{digits}f}, {hi:.{digits}f}]"
"""## 12. Logging"""
EVAL_LOG_DIR = Path("./eval_logs")
EVAL_LOG_DIR.mkdir(exist_ok=True)
def log_eval_row(experiment_id, passage_id, variant, retrieved_sources,
generation_result, judge_scores, extra=None):
row = {
"experiment_id": experiment_id,
"passage_id": passage_id,
"variant": variant,
"model": GENERATOR_MODEL,
"n_retrieved": len(retrieved_sources),
"retrieved_chunk_ids": ";".join(
str(r.get("chunk_id") or r.get("concept_id") or "?") for r in retrieved_sources
),
"guard_triggered": int(generation_result.get("guard_triggered", False)),
"abstained": int(generation_result.get("abstained", False)),
"answer_chars": len(generation_result.get("answer", "")),
"generated_text": generation_result.get("answer", ""),
"correctness": judge_scores.get("correctness", {}).get("score"),
"evidence_support": judge_scores.get("evidence_support", {}).get("score"),
"citation_faithfulness": judge_scores.get("citation_faithfulness", {}).get("score"),
"semantic_fail_rate": generation_result.get("semantic_fail_rate", np.nan),
}
if extra:
row.update(extra)
path = EVAL_LOG_DIR / f"{experiment_id}.csv"
pd.DataFrame([row]).to_csv(
path, mode="a", header=not path.exists(), index=False
)
return row
def load_or_run_experiment(experiment_id, runner_fn):
local_path = EVAL_LOG_DIR / f"{experiment_id}.csv"
if RUN_EXPERIMENTS:
if local_path.exists():
local_path.unlink()
print(f"running {experiment_id} from scratch...")
return runner_fn()
url = f"{REPO_RAW_BASE}/eval_logs/{experiment_id}.csv"
try:
df = pd.read_csv(url)
print(f"loaded {experiment_id} from repo cache: {len(df)} rows")
df.to_csv(local_path, index=False)
return df
except Exception:
pass
if local_path.exists():
df = pd.read_csv(local_path)
print(f"loaded {experiment_id} from local cache: {len(df)} rows")
return df
print(f"⚠ no cached results for {experiment_id}. Set RUN_EXPERIMENTS=True to generate.")
return None
"""## 13. Judge calibration
"""
RUN_CALIBRATION = False # HERE
def calibrate_judge(n_items=5):
items = [x for x in validation_set["items"]]
sample = items[:n_items]
diffs = {"correctness": [], "evidence_support": [], "citation_faithfulness": []}
for item in sample:
c, p = retrieve_for_variant(item["passage"], "hybrid")
result = generate_with_citation_guard(item["passage"], c, p, do_semantic_check=False)
explanation = result["answer"]
context_text = _format_context_block(c, p)
print("=" * 70)
print(f"ITEM {item['id']}")
print(f"PASSAGE: {item['passage'][:300]}")
print(f"\nREFERENCE: {item['reference_explanation']}")
print(f"\nSYSTEM EXPLANATION:\n{explanation}")
print("\nScore each metric 0/1/2 (0=bad, 1=partial, 2=good):")
try:
human = {
"correctness": int(input(" correctness: ")),
"evidence_support": int(input(" evidence_support: ")),
"citation_faithfulness": int(input(" citation_faithfulness: ")),
}
except (ValueError, EOFError):
print("aborted")
return None
all_scores = score_all_metrics(
item["passage"], item["reference_explanation"], context_text, explanation
)
scores_clean = {k: all_scores[k]["score"] for k in all_scores}
for k in diffs:
if scores_clean[k] is not None:
diffs[k].append(abs(human[k] - scores_clean[k]))
print("\n=== CALIBRATION RESULTS ===")
for k, vals in diffs.items():
if vals:
mad = sum(vals) / len(vals)
flag = " ⚠ DISAGREES" if mad > 0.5 else " ok"
print(f" {k}: mean abs diff = {mad:.2f}{flag}")
return diffs
if RUN_CALIBRATION:
calibrate_judge(n_items=5)
else:
print("calibration skipped (RUN_CALIBRATION=False)")
print("Last calibration: correctness MAD=0.60 (DISAGREES), evidence MAD=0.40, citation MAD=0.20")
"""## 14. Experiment A — retrieval ablation
**Question.** Does RAG help, and does the regex tier earn its place?
**Hypothesis.** All RAG variants will beat the no-RAG baseline on claim_support_rate and evidence_support. Hybrid will beat either single-tier variant.
**Variable changed.** Retrieval method ∈ {no_rag, embedding_only, regex_only, hybrid}. Everything else held constant.
"""
def run_experiment_A():
items = [x for x in validation_set["items"]]
variants = ["no_rag", "embedding_only", "regex_only", "hybrid"]
total_runs = len(items) * len(variants)
print(f"running experiment A: {len(items)} items × {len(variants)} variants = {total_runs} runs")
for i, item in enumerate(items):
for variant in variants:
try:
c, p = retrieve_for_variant(item["passage"], variant)
retrieved = c + p
context_text = _format_context_block(c, p)
result = generate_with_citation_guard(
item["passage"], c, p,
allow_no_context_bypass=(variant == "no_rag"),
do_semantic_check=False,
)
scores = score_all_metrics(
item["passage"], item["reference_explanation"],
context_text, result["answer"],
)
cb = claim_based_faithfulness(
result["answer"], [r["content"] for r in retrieved],
)
rp, rr = precision_recall_at_k(
[r["content"] for r in retrieved], item["key_terms"], k=3,
)
log_eval_row(
"experiment_A", item["id"], variant,
retrieved, result, scores,
extra={
"category": item["category"],
"claim_support_rate": cb["support_rate"],
"claim_contradiction_rate": cb["contradiction_rate"],
"claim_unsupported_rate": cb["unsupported_rate"],
"n_claims": cb["n_claims"],
"retrieval_precision_at_3": rp,
"retrieval_recall_at_3": rr,
}
)
except Exception as e:
print(f" ERROR {item['id']}/{variant}: {e}")
print(f" done {item['id']} ({i+1}/{len(items)})")
return pd.read_csv(EVAL_LOG_DIR / "experiment_A.csv")
experiment_A_df = load_or_run_experiment("experiment_A", run_experiment_A)
def analyze_experiment_A():
df = pd.read_csv(EVAL_LOG_DIR / "experiment_A.csv")
metric_cols = ["correctness", "evidence_support", "citation_faithfulness",
"claim_support_rate", "retrieval_recall_at_3", "abstained"]
print("=" * 70)
print("OVERALL means with 95% bootstrap CIs")
print("=" * 70)
for variant in ["no_rag", "embedding_only", "regex_only", "hybrid"]:
sub = df[df.variant == variant]
print(f"\n{variant}")
for m in metric_cols:
if m in sub.columns:
print(f" {m:28s} {format_ci(sub[m].values)}")
print("\n" + "=" * 70)
print("HEADLINE METRIC: claim_support_rate (correctness saturates — see report)")
print("=" * 70)
for variant in ["no_rag", "embedding_only", "regex_only", "hybrid"]:
sub = df[df.variant == variant]
if "claim_support_rate" in sub.columns:
print(f" {variant:18s} {format_ci(sub['claim_support_rate'].values)}")
return df
def plot_experiment_A():
df = pd.read_csv(EVAL_LOG_DIR / "experiment_A.csv")
variant_order = ["no_rag", "embedding_only", "regex_only", "hybrid"]
colors = ["#888", "#4c72b0", "#dd8452", "#55a868"]
fig, axes = plt.subplots(1, 3, figsize=(16, 5))
means, los, his = [], [], []
for v in variant_order:
sub = df[df.variant == v]
if "claim_support_rate" in sub.columns:
m, lo, hi = bootstrap_ci(sub["claim_support_rate"].values)
else:
m, lo, hi = 0, 0, 0
means.append(m); los.append(m - lo); his.append(hi - m)
axes[0].bar(variant_order, means, yerr=[los, his], color=colors, capsize=5)
axes[0].set_title("Claim support rate (headline)")
axes[0].set_ylabel("Fraction of claims supported")
axes[0].set_ylim(0, 1)
axes[0].tick_params(axis="x", rotation=20)
means, los, his = [], [], []
for v in variant_order:
sub = df[df.variant == v]
if "retrieval_recall_at_3" in sub.columns:
m, lo, hi = bootstrap_ci(sub["retrieval_recall_at_3"].values)
else:
m, lo, hi = 0, 0, 0
means.append(m); los.append(m - lo); his.append(hi - m)
axes[1].bar(variant_order, means, yerr=[los, his], color=colors, capsize=5)
axes[1].set_title("Retrieval recall@3")
axes[1].set_ylabel("Fraction of gold key_terms covered")
axes[1].set_ylim(0, 1)
axes[1].tick_params(axis="x", rotation=20)
abs_by_var = df.groupby("variant")["abstained"].mean().reindex(variant_order)
axes[2].bar(variant_order, abs_by_var.values, color=colors)
axes[2].set_title("Abstention rate")
axes[2].set_ylabel("Fraction of items guard triggered")
axes[2].set_ylim(0, 1)
axes[2].tick_params(axis="x", rotation=20)
plt.tight_layout()
plt.show()
if experiment_A_df is not None:
analyze_experiment_A()
plot_experiment_A()
"""### Release gate"""
def release_gate_A(variant="hybrid"):
df = pd.read_csv(EVAL_LOG_DIR / "experiment_A.csv")
sub = df[df.variant == variant]
thresholds = {
"claim_support_rate": 0.70, # primary
"evidence_support": 1.40,
"citation_faithfulness": 1.40,
"retrieval_recall_at_3": 0.60,
"abstained": 0.30,
}
lower_is_better = {"abstained"}
agg = {k: float(np.nanmean(sub[k].values)) for k in thresholds if k in sub.columns}
print(f"Release gate for variant: {variant}")
print("=" * 60)
all_pass = True
for k, t in thresholds.items():
if k not in agg: continue
v = agg[k]
ok = (v <= t) if k in lower_is_better else (v >= t)
direction = "≤" if k in lower_is_better else "≥"
status = "PASS" if ok else "FAIL"
print(f" {k:28s} {v:.3f} (need {direction} {t}) {status}")
all_pass = all_pass and ok
print(f"\nFINAL: {'PASS' if all_pass else 'FAIL'}")
return all_pass
if experiment_A_df is not None:
release_gate_A(variant="hybrid")
print()
release_gate_A(variant="regex_only")
"""## 16. Experiment B — top-k sweep
**Question:** How does the number of retrieved sources (top-k) affect answer correctness?
**Hypothesis:** Performance peaks somewhere in the middle. k=1 misses context; large k dilutes the prompt with irrelevant chunks.
**Variable changed:** `top_k ∈ {1, 3, 5, 7}`, applied to both retrieval tiers. Hybrid retrieval; everything else held constant.
"""
def run_experiment_B(top_k_values=(1, 3, 5, 7)):
items = [x for x in validation_set["items"]]
print(f"running experiment B: {len(items)} items × {len(top_k_values)} top-k values "
f"= {len(items) * len(top_k_values)} runs")
for k in top_k_values:
time.sleep(1.5)
print(f"\n--- top_k = {k} ---")
for item in items:
try:
c, p = retrieve_for_variant(
item["passage"], "hybrid", n_concepts=k, n_papers=k,
)
retrieved = c + p
context_text = _format_context_block(c, p)
result = generate_with_citation_guard(
item["passage"], c, p, do_semantic_check=False
)
scores = score_all_metrics(
item["passage"], item["reference_explanation"],
context_text, result["answer"]
)
avg_dist = (float(np.mean([r["distance"] for r in retrieved if r["distance"] > 0]))
if retrieved else None)
log_eval_row(
"experiment_B", item["id"], f"topk_{k}",
retrieved, result, scores,
extra={
"category": item["category"],
"top_k": k,
"n_retrieved_total": len(retrieved),
"avg_distance": avg_dist,
}
)
except Exception as e:
print(f" ERROR {item['id']}: {e}")
return pd.read_csv(EVAL_LOG_DIR / "experiment_B.csv")
def analyze_experiment_B():
df = pd.read_csv(EVAL_LOG_DIR / "experiment_B.csv")
print("mean correctness by top-k (with 95% CI):")
for k in sorted(df["top_k"].unique()):
sub = df[df.top_k == k]
print(f" top_k={k} "
f"correctness={format_ci(sub['correctness'].values)} "
f"evidence={format_ci(sub['evidence_support'].values)} "
f"avg_n_retrieved={sub['n_retrieved_total'].mean():.1f}")
return df
def plot_experiment_B():
df = pd.read_csv(EVAL_LOG_DIR / "experiment_B.csv")
ks = sorted(df["top_k"].unique())
means, los, his = [], [], []
for k in ks:
m, lo, hi = bootstrap_ci(df[df.top_k == k]["correctness"].values)
means.append(m); los.append(m - lo); his.append(hi - m)
fig, ax = plt.subplots(figsize=(8, 5))
ax.errorbar(ks, means, yerr=[los, his], marker="o", linewidth=2, capsize=5)
ax.set_xlabel("top-k (per retrieval tier)")
ax.set_ylabel("Mean correctness (0-2)")
ax.set_title("Experiment B — Correctness vs top-k (95% CI)")
ax.set_xticks(ks)
ax.set_ylim(0, 2)
ax.grid(alpha=0.3)
plt.tight_layout()
plt.show()
experiment_B_df = load_or_run_experiment("experiment_B", run_experiment_B)
if experiment_B_df is not None:
analyze_experiment_B()
plot_experiment_B()
"""## 17. Experiment C — confidence threshold tuning
**Question.** Where should the low-confidence threshold sit so that warnings correlate with wrong answers?
**Hypothesis.** The default 1.3 threshold was a guess. The F1-maximizing threshold is probably lower.
**Variable changed.** Threshold ∈ [0.6, 1.6] by 0.1.
"""
def run_experiment_C():
items = [x for x in validation_set["items"]]
for item in items:
try:
time.sleep(1.5)
c, p = retrieve_for_variant(item["passage"], "hybrid")
all_dists = [r["distance"] for r in (c + p) if r["distance"] > 0]
best_dist = float(min(all_dists)) if all_dists else 999.0
context_text = _format_context_block(c, p)
result = generate_with_citation_guard(
item["passage"], c, p, do_semantic_check=False
)
scores = score_all_metrics(
item["passage"], item["reference_explanation"], context_text, result["answer"]
)
corr = scores["correctness"]["score"]
log_eval_row(
"experiment_C", item["id"], "default_system",
c + p, result, scores,
extra={
"category": item["category"],
"best_distance": best_dist,
"correctness_raw": corr,
}
)
except Exception as e:
print(f" ERROR {item['id']}: {e}")
return pd.read_csv(EVAL_LOG_DIR / "experiment_C.csv")
def analyze_experiment_C():
df = pd.read_csv(EVAL_LOG_DIR / "experiment_C.csv")
if "correctness_raw" not in df.columns:
df["correctness_raw"] = df["correctness"]
df = df.dropna(subset=["best_distance", "correctness_raw"])
strict_pos = (df["correctness_raw"] < 1.0).sum()
if strict_pos > 0:
df["is_wrong"] = (df["correctness_raw"] < 1.0).astype(int)
wrongness_def = "correctness < 1.0 (strict)"
else:
df["is_wrong"] = (df["correctness_raw"] < 2.0).astype(int)
wrongness_def = "correctness < 2.0 (saturation fallback)"
print(f"using wrongness definition: {wrongness_def}")
print(f" positive class size: {df['is_wrong'].sum()}/{len(df)}")
if df["is_wrong"].sum() == 0:
print("⚠ WARNING: no wrong answers in eval set. Tuning is meaningless on this data.")
return None, None, None
thresholds = [round(0.6 + 0.1 * i, 2) for i in range(11)]
rows = []
for t in thresholds:
warns = df["best_distance"] > t
wrong = df["is_wrong"] == 1
tp = int(((warns) & (wrong)).sum())
fp = int(((warns) & (~wrong)).sum())
fn = int(((~warns) & (wrong)).sum())
tn = int(((~warns) & (~wrong)).sum())
precision = tp / (tp + fp) if (tp + fp) > 0 else 1.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = 2*precision*recall / (precision + recall) if (precision + recall) > 0 else 0.0
tpr = recall
fpr = fp / (fp + tn) if (fp + tn) > 0 else 0.0
rows.append({"threshold": t, "tp": tp, "fp": fp, "fn": fn, "tn": tn,
"refusal_precision": round(precision, 3),
"refusal_recall": round(recall, 3),
"f1": round(f1, 3),
"tpr": round(tpr, 3), "fpr": round(fpr, 3)})
sweep = pd.DataFrame(rows)
print(sweep)
best = sweep.loc[sweep["f1"].idxmax()]
print(f"\nF1-maximizing threshold: {best['threshold']} (F1={best['f1']})")
print(f" refusal precision: {best['refusal_precision']}")
print(f" refusal recall: {best['refusal_recall']}")
s = sweep.sort_values("fpr")
auc = 0.0
for i in range(1, len(s)):
auc += (s.iloc[i]["fpr"] - s.iloc[i-1]["fpr"]) * (s.iloc[i]["tpr"] + s.iloc[i-1]["tpr"]) / 2
print(f"approx ROC AUC: {auc:.3f}")
return sweep, best, auc
def plot_experiment_C():
out = analyze_experiment_C()
if out is None or out[0] is None:
return
sweep, best, auc = out
fig, ax = plt.subplots(figsize=(7, 7))
s = sweep.sort_values("fpr")
ax.plot(s["fpr"], s["tpr"], marker="o", linewidth=2, label=f"ROC (AUC≈{auc:.3f})")
ax.plot([0, 1], [0, 1], "--", color="gray", alpha=0.5, label="chance")
ax.scatter([best["fpr"]], [best["tpr"]], s=200, color="red", zorder=5,
label=f"best F1 @ threshold={best['threshold']}")
ax.set_xlabel("False positive rate")
ax.set_ylabel("True positive rate (refusal recall)")
ax.set_title("Experiment C — Abstention threshold ROC")
ax.set_xlim(-0.05, 1.05); ax.set_ylim(-0.05, 1.05)
ax.legend(loc="lower right")
ax.grid(alpha=0.3)
plt.tight_layout()
plt.show()
experiment_C_df = load_or_run_experiment("experiment_C", run_experiment_C)
if experiment_C_df is not None:
plot_experiment_C()
DEFAULT_CONFIDENCE_THRESHOLD = 1.3
try:
if experiment_C_df is not None:
out = analyze_experiment_C()
if out and out[1] is not None:
_, best, _ = out
TUNED_CONFIDENCE_THRESHOLD = float(best["threshold"])
else:
TUNED_CONFIDENCE_THRESHOLD = DEFAULT_CONFIDENCE_THRESHOLD
else:
TUNED_CONFIDENCE_THRESHOLD = DEFAULT_CONFIDENCE_THRESHOLD
except Exception as e:
print(f"falling back to default threshold ({e})")
TUNED_CONFIDENCE_THRESHOLD = DEFAULT_CONFIDENCE_THRESHOLD
print(f"TUNED_CONFIDENCE_THRESHOLD = {TUNED_CONFIDENCE_THRESHOLD}")
"""## 18. Main pipeline (with citations + tuned threshold + semantic check)"""
def check_input_quality(text):
if len(text.strip()) < 20:
return False, "That's pretty short — try pasting a full sentence or paragraph from a paper."
if len(text.strip()) > 3000:
return False, "That's a lot of text. Try pasting just 1-2 paragraphs at a time."
if len(text.split()) < 5:
return False, "Try a longer passage — at least a full sentence from a paper."
return True, "ok"
def assess_retrieval_confidence(concept_results, paper_results, threshold=None):
threshold = threshold if threshold is not None else TUNED_CONFIDENCE_THRESHOLD
dists = [r["distance"] for r in (concept_results + paper_results) if r["distance"] > 0]
if not dists:
return "low", "I couldn't find any relevant context in my knowledge base."
best = min(dists)
if best < 0.8:
return "high", ""
elif best < threshold:
return "medium", ("Note: my knowledge base has some related material, but the match isn't perfect. "
"Double-check against the paper's own definitions.")
else:
return "low", "Heads up: the concepts in this passage don't match well with my current knowledge base."
SCOPE_DISCLAIMER = (
"---\n"
"*This tool helps you understand papers; it doesn't replace them. "
"Every factual sentence above is cited to a specific retrieved source. "
"⚠️ marks indicate the semantic guard flagged that sentence as not fully supported by its citation. "
"Always check the original paper.*"
)
def scimplify(passage, variant="hybrid"):
is_ok, msg = check_input_quality(passage)
if not is_ok:
return msg
c, p = retrieve_for_variant(passage, variant)
confidence, warning = assess_retrieval_confidence(c, p)
result = generate_with_citation_guard(passage, c, p)
parts = []
if result["guard_triggered"]:
which = "semantic" if any("semantic" in i for i in result.get("issues", [])) else "lexical"
parts.append(f"⚠️ The {which} citation guard triggered. Returning abstention rather than a potentially ungrounded answer.")
if result.get("issues"):
parts.append(f"\n*Reason: {'; '.join(result['issues'])}*")
parts.append(f"\n{result['answer']}")
parts.append(f"\n{SCOPE_DISCLAIMER}")
return "\n".join(parts)
if result["answer"].strip() == ABSTAIN_MESSAGE or result["abstained"]:
parts = [result["answer"], SCOPE_DISCLAIMER]
return "\n".join(parts)
if confidence == "low":
parts.append(f"⚠️ {warning}\n")
elif confidence == "medium":
parts.append(f"ℹ️ {warning}\n")
parts.append(result["answer"])
# show retrieved sources
concept_names = [r["concept_name"] for r in c if "concept_name" in r]
if concept_names:
parts.append(f"\n\n**Retrieved concepts:** {', '.join(concept_names)}")
if p:
sources = sorted(set(r["source_name"] for r in p))
parts.append(f"**Paper sources:** {', '.join(sources)}")
# surface semantic check stats if any sentences were checked
findings = result.get("semantic_findings", [])
if findings:
n_total = len(findings)
n_unsupported = sum(1 for f in findings if f["label"] in ("contradicted", "insufficient"))
if n_unsupported > 0:
parts.append(f"\n*Semantic guard: {n_unsupported}/{n_total} cited sentences flagged as not fully supported.*")
else:
parts.append(f"\n*Semantic guard: all {n_total} cited sentences supported by their citations ✓*")
parts.append(f"\n{SCOPE_DISCLAIMER}")
return "\n".join(parts)
"""## 19. Gradio UI
"""
def add_pdf_to_kb(pdf_file, source_name, source_type):
if pdf_file is None:
return "Please upload a PDF file."
if not source_name.strip():
return "Please provide a name for this source."
try:
text = extract_text_from_pdf(pdf_file)
chunks = chunk_text(text)
base = source_name.strip().replace(" ", "_")
ids = [f"user_{base}::c{i}" for i in range(len(chunks))]
metas = [{
"source_name": source_name.strip(),
"source_type": source_type,
"chunk_id": ids[i],
} for i in range(len(chunks))]
if chunks:
papers_collection.add(documents=chunks, ids=ids, metadatas=metas)
return f"Added {len(chunks)} chunks. Total: {papers_collection.count()}"
except Exception as e:
return f"Error: {e}"
def pull_from_arxiv_ui(query, max_results):
"""Gradio handler for arXiv ingestion."""
try:
max_results = int(max_results)
if max_results < 1 or max_results > 25:
return "Please pick a max_results between 1 and 25."
summary = ingest_from_arxiv(query=query, max_results=max_results, verbose=False)
msg = (
f"✅ Ingested {summary['n_papers']} new paper(s), "
f"{summary['n_chunks']} chunks. "
f"Skipped {summary['n_skipped']} duplicates. "
f"Total in KB: {summary['total_in_kb']} chunks."
)
if summary["errors"]:
msg += f"\n\n⚠️ Errors: {'; '.join(summary['errors'][:3])}"
return msg
except Exception as e:
return f"Error: {e}"
def get_kb_status():
n_concepts = concepts_collection.count()
n_papers = papers_collection.count()
status = f"**Concept definitions:** {n_concepts}\n\n**Paper chunks:** {n_papers}\n"
if n_papers > 0:
metas = papers_collection.get()["metadatas"]
sources = Counter(m["source_name"] for m in metas)
status += "\n**Ingested sources:**\n"
for name, count in sources.most_common():
status += f"- {name} — {count} chunks\n"
return status
DEMO_CLEAN = "The Diels-Alder reaction is a [4+2] cycloaddition between a conjugated diene and a dienophile, producing a six-membered ring with up to four new stereocenters. The reaction proceeds through a concerted, suprafacial transition state and is highly stereospecific: cis-dienophiles yield cis-substituted cyclohexenes. Electron-withdrawing groups on the dienophile dramatically accelerate the reaction."
DEMO_PAPER = "Multi-region neural population dynamics in the brain have been studied using techniques like LFADS to model the latent factors driving observed activity across regions."
DEMO_ABSTAIN = "Laminated pastry dough is created by repeatedly folding butter into flour-water dough, producing alternating layers that puff up during baking as steam expands between them. Croissants are the canonical example."
with gr.Blocks(title="Scimplify") as app:
gr.Markdown("# Scimplify — NeuroAI Paper Simplifier")
gr.Markdown(
"Paste a NeuroAI paragraph; get a plain-language explanation with citations. "
"Every factual sentence is grounded in a retrieved source. The lexical guard rejects "
"invented citation IDs, and the semantic guard verifies that each cited chunk actually "
"supports the claim. If neither passes, the system abstains rather than hallucinate."
)
with gr.Tab("Explain Passage"):
with gr.Row():
with gr.Column(scale=1):
inp = gr.Textbox(label="Passage", lines=8,
placeholder="Paste a paragraph from a paper...")
btn = gr.Button("Explain", variant="primary")
gr.Examples(
examples=[
[DEMO_CLEAN],
[DEMO_PAPER],
[DEMO_ABSTAIN],
],
inputs=[inp],
label="Demo passages (clean / paper-chunk / out-of-scope)",
)
with gr.Column(scale=2):
out = gr.Markdown(label="Explanation")
btn.click(fn=lambda x: scimplify(x), inputs=[inp], outputs=[out])
with gr.Tab("Add Papers (PDF)"):
pdf_in = gr.File(label="PDF", file_types=[".pdf"])
name_in = gr.Textbox(label="Source name")
type_in = gr.Radio(["paper", "article", "review"], label="Type", value="paper")
add_btn = gr.Button("Add to knowledge base")
add_out = gr.Textbox(label="Status")
add_btn.click(fn=add_pdf_to_kb, inputs=[pdf_in, name_in, type_in], outputs=[add_out])
with gr.Tab("Pull from arXiv"):
gr.Markdown(
"Fetch recent NeuroAI papers from arXiv directly. "
"Skips papers already in the knowledge base (matched by arxiv_id)."
)
arxiv_query = gr.Textbox(
label="arXiv query",
value="NeuroAI",
placeholder="e.g. NeuroAI, brain-inspired deep learning, neural population dynamics",
)
arxiv_n = gr.Slider(label="Max papers", minimum=1, maximum=20, value=5, step=1)
arxiv_btn = gr.Button("Pull from arXiv", variant="primary")
arxiv_out = gr.Markdown()
arxiv_btn.click(fn=pull_from_arxiv_ui, inputs=[arxiv_query, arxiv_n], outputs=[arxiv_out])
with gr.Tab("Knowledge Base"):
status_out = gr.Markdown(value=get_kb_status())
refresh_btn = gr.Button("Refresh")
refresh_btn.click(fn=get_kb_status, outputs=[status_out])
app.launch(share=True) |