Spaces:
Running
Running
File size: 60,189 Bytes
157fe17 | 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 | {
"nbformat": 4,
"nbformat_minor": 5,
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {"name": "ipython", "version": 3},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"version": "3.10.0"
}
},
"cells": [
{
"cell_type": "markdown",
"id": "md-title",
"metadata": {},
"source": [
"# RAGBench Knowledge Base Builder — v2\n",
"\n",
"Builds **FAISS + BM25 + chunk/doc artefacts** for every\n",
"`(dataset × embedding_model × chunk_config)` combination and uploads each\n",
"to a Hugging Face dataset repo so the Gradio app can download on demand.\n",
"\n",
"## Design decisions\n",
"\n",
"| Decision | Choice | Rationale |\n",
"|---|---|---|\n",
"| Corpus scope (v1) | `test` split only | 12 datasets × models × chunk configs is significant compute; test split is sufficient for comparative evaluation |\n",
"| Evaluation setting | **Closed retrieval** | Every test question's relevant document is guaranteed in the index; metrics measure *relative quality* across configs, not open-corpus absolute recall |\n",
"| Future expansion | `all_splits` rebuild | Change `corpus_scope` → new `index_id` automatically; old indices untouched |\n",
"| Existing indices | Deleted + rebuilt | Prior indices had an incomplete cache key and silent truncation (chunk_size > model max_seq_length) |\n",
"\n",
"## Key improvements over v1\n",
"\n",
"1. **Parameterised corpus scope** — `corpus_scope` field drives the loader; upgrading to `all_splits` later needs only one field change.\n",
"2. **Pre-build validation** — chunk sizes are checked against every model's effective `max_seq_length` before any encoding starts.\n",
"3. **Query / passage prefixes** — BGE and other asymmetric models require different text formats for passages vs queries; both are stored in every combo and in `meta.json`.\n",
"4. **Chunking strategy dispatch** — `chunking_strategy` now selects a real implementation (`passthrough`, `recursive`); unknown strategies raise immediately.\n",
"5. **Complete cache key** — includes schema version, corpus scope, model revision, prefixes, and strategy; two different indices can never share an `index_id`.\n",
"6. **Improved BM25 tokenisation** — regex word splitting separates punctuation from legal terms (e.g. `clause,` → `clause`).\n",
"7. **Richer `meta.json`** — stores every field needed to reproduce or audit the index.\n",
"\n",
"## Colab workflow\n",
"\n",
"1. Set `HF_TOKEN` in the Colab secrets panel (🔑 left sidebar).\n",
"2. Edit `HF_REPO` in **Cell 2**.\n",
"3. Run **Cells 1–5** once per session (install + auth + helpers).\n",
"4. Run **Cell 6** (validation) — confirms all combos are safe before any compute.\n",
"5. Run **Cell 7** to see which combos are done vs pending.\n",
"6. Set `COMBO_IDX` in **Cell 8** and run **Cells 8–12** to build and upload one combo.\n",
"7. Repeat step 6 in a fresh session for the next pending combo.\n",
"\n",
"Or run **Cell 13** to build all pending combos automatically in one session."
]
},
{
"cell_type": "markdown",
"id": "md-install",
"metadata": {},
"source": [
"## Cell 1 — Install dependencies"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-install",
"metadata": {},
"outputs": [],
"source": [
"!pip install -q datasets transformers sentence-transformers faiss-cpu rank_bm25 langchain-text-splitters huggingface_hub pyarrow pandas numpy"
]
},
{
"cell_type": "markdown",
"id": "md-config",
"metadata": {},
"source": [
"## Cell 2 — Configuration\n",
"\n",
"### Corpus scope\n",
"\n",
"RAGBench provides train / validation / test splits of *question–answer–context triples*,\n",
"not of a separate document corpus. In a production RAG system all documents are indexed;\n",
"queries are evaluated against that one index regardless of which split they came from.\n",
"\n",
"For this first build (`corpus_scope = \"test\"`) we index only the documents attached to\n",
"test-split examples. This creates a **closed retrieval** setting: every test question's\n",
"relevant document is guaranteed to be in the index. Retrieval metrics (Hit@k, nDCG@k)\n",
"therefore measure *relative quality* across configurations, not open-corpus absolute recall.\n",
"This is documented explicitly in every `meta.json` artifact.\n",
"\n",
"**Upgrade path:** change `corpus_scope` to `\"all_splits\"` in any combo. The updated\n",
"`make_cache_key()` will produce a different `index_id`; old test-scope indices stay\n",
"untouched in the HF repo and both are listed in `index_config.json`.\n",
"\n",
"### Combo schema\n",
"\n",
"| Field | Purpose |\n",
"|---|---|\n",
"| `corpus_scope` | Which splits to load for the document corpus (`test` / `train` / `all_splits`) |\n",
"| `embedding_model` | HuggingFace model ID |\n",
"| `model_revision` | Git ref — pins the model checkpoint for reproducibility |\n",
"| `query_prefix` | Instruction prepended to queries **at runtime** by the retriever |\n",
"| `passage_prefix` | Instruction prepended to passages **during indexing** |\n",
"| `chunking_strategy` | `passthrough`, `recursive`, or `parent_child` |\n",
"| `chunk_size` | Target chunk size in tokens (0 = not applicable for passthrough) |\n",
"| `chunk_overlap` | Overlap between adjacent chunks in tokens |\n",
"\n",
"### Asymmetric model prefixes\n",
"\n",
"Some embedding models use asymmetric formatting:\n",
"- **BGE v1.5** — short instruction on retrieval queries; no prefix on passages.\n",
"- **E5** — `query: ` on queries, `passage: ` on passages.\n",
"- **GTE-ModernBERT / MiniLM / MPNet** — symmetric raw-text embedding (empty prefixes).\n",
"\n",
"The `passage_prefix` is applied during index construction (this notebook).\n",
"The `query_prefix` must be applied in the app's `HybridRetriever.dense_search()`\n",
"and read from `meta.json` — **both must change together**.\n",
"\n",
"### Chunking strategies\n",
"\n",
"- **`passthrough`** — each source passage becomes one chunk. Best for short retrieved\n",
" passages (e.g. CovidQA ~100 words) that already fit within any model's limit.\n",
"- **`recursive`** — tokenizer-aware `RecursiveCharacterTextSplitter`. `chunk_size` is in\n",
" the model's own token space and validated against `model.max_seq_length` before building.\n",
"- **`parent_child`** — small retrieval children with a larger parent stored for generation.\n",
" Not yet implemented; requires retrieval + generation pipeline changes."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-config",
"metadata": {},
"outputs": [],
"source": [
"# ── Schema version ─────────────────────────────────────────────────────────────\n",
"# Increment when any build-logic change would produce a different index from the\n",
"# same combo settings. Forces all affected combos to a new index_id.\n",
"SCHEMA_VERSION = 2\n",
"\n",
"HF_REPO = \"your-username/rag-capstone-indices\" # <── change to your HF repo\n",
"\n",
"SPLIT_MAP = {\n",
" \"test\": [\"test\"],\n",
" \"train\": [\"train\"],\n",
" \"all_splits\": [\"train\", \"validation\", \"test\"],\n",
"}\n",
"\n",
"ARTIFACT_FILES = [\n",
" \"faiss.index\", \"embeddings.npy\", \"bm25.pkl\",\n",
" \"chunks.parquet\", \"docs.parquet\", \"meta.json\",\n",
"]\n",
"\n",
"# ── CovidQA combos — biomedical domain ────────────────────────────────────────\n",
"# Source passages are already short (~100 words from CORD-19 retrieval).\n",
"# Strategy: passthrough — each passage becomes exactly one chunk.\n",
"# chunk_size=0 / chunk_overlap=0 are sentinels meaning \"not applicable\".\n",
"COVIDQA_COMBOS = [\n",
" {\n",
" \"dataset\": \"covidqa\", \"domain\": \"biomedical\",\n",
" \"display_name\": \"CovidQA (Biomedical)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"sentence-transformers/all-MiniLM-L6-v2\",\n",
" \"embedding_display_name\": \"MiniLM L6 v2\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"passthrough\",\n",
" \"chunk_size\": 0, \"chunk_overlap\": 0,\n",
" },\n",
" {\n",
" \"dataset\": \"covidqa\", \"domain\": \"biomedical\",\n",
" \"display_name\": \"CovidQA (Biomedical)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"BAAI/bge-base-en-v1.5\",\n",
" \"embedding_display_name\": \"BGE base en v1.5\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"Represent this sentence for searching relevant passages: \",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"passthrough\",\n",
" \"chunk_size\": 0, \"chunk_overlap\": 0,\n",
" },\n",
" {\n",
" \"dataset\": \"covidqa\", \"domain\": \"biomedical\",\n",
" \"display_name\": \"CovidQA (Biomedical)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"Alibaba-NLP/gte-modernbert-base\",\n",
" \"embedding_display_name\": \"GTE ModernBERT base\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"passthrough\",\n",
" \"chunk_size\": 0, \"chunk_overlap\": 0,\n",
" },\n",
" {\n",
" \"dataset\": \"covidqa\", \"domain\": \"biomedical\",\n",
" \"display_name\": \"CovidQA (Biomedical)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"sentence-transformers/all-mpnet-base-v2\",\n",
" \"embedding_display_name\": \"MPNet base v2\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"passthrough\",\n",
" \"chunk_size\": 0, \"chunk_overlap\": 0,\n",
" },\n",
"]\n",
"\n",
"# ── CUAD combos — legal domain ─────────────────────────────────────────────────\n",
"# Full commercial contracts (1–100 pages). Recursive chunking required.\n",
"# All chunk sizes stay within each model's max_seq_length (validated in Cell 6).\n",
"CUAD_COMBOS = [\n",
" # Tier 1: fast baseline — MiniLM (max_seq_length=256)\n",
" {\n",
" \"dataset\": \"cuad\", \"domain\": \"legal\",\n",
" \"display_name\": \"CUAD (Legal Contracts)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"sentence-transformers/all-MiniLM-L6-v2\",\n",
" \"embedding_display_name\": \"MiniLM L6 v2\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"recursive\",\n",
" \"chunk_size\": 256, \"chunk_overlap\": 32,\n",
" },\n",
" # Tier 2: cross-domain baseline — MPNet (max_seq_length=384)\n",
" {\n",
" \"dataset\": \"cuad\", \"domain\": \"legal\",\n",
" \"display_name\": \"CUAD (Legal Contracts)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"sentence-transformers/all-mpnet-base-v2\",\n",
" \"embedding_display_name\": \"MPNet base v2\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"recursive\",\n",
" \"chunk_size\": 384, \"chunk_overlap\": 64,\n",
" },\n",
" # Tier 3: strong dense retrieval — BGE-large (max_seq_length=512, asymmetric)\n",
" {\n",
" \"dataset\": \"cuad\", \"domain\": \"legal\",\n",
" \"display_name\": \"CUAD (Legal Contracts)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"BAAI/bge-large-en-v1.5\",\n",
" \"embedding_display_name\": \"BGE large en v1.5\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"Represent this sentence for searching relevant passages: \",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"recursive\",\n",
" \"chunk_size\": 512, \"chunk_overlap\": 96,\n",
" },\n",
" # Tier 4: modern long-context — GTE-ModernBERT (max_seq_length=8192), 512 tokens\n",
" {\n",
" \"dataset\": \"cuad\", \"domain\": \"legal\",\n",
" \"display_name\": \"CUAD (Legal Contracts)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"Alibaba-NLP/gte-modernbert-base\",\n",
" \"embedding_display_name\": \"GTE ModernBERT base\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"recursive\",\n",
" \"chunk_size\": 512, \"chunk_overlap\": 96,\n",
" },\n",
" # Tier 4b: long-clause experiment — GTE-ModernBERT, 1024 tokens\n",
" {\n",
" \"dataset\": \"cuad\", \"domain\": \"legal\",\n",
" \"display_name\": \"CUAD (Legal Contracts)\",\n",
" \"corpus_scope\": \"test\",\n",
" \"embedding_model\": \"Alibaba-NLP/gte-modernbert-base\",\n",
" \"embedding_display_name\": \"GTE ModernBERT base (long clause)\",\n",
" \"model_revision\": \"main\",\n",
" \"query_prefix\": \"\",\n",
" \"passage_prefix\": \"\",\n",
" \"chunking_strategy\": \"recursive\",\n",
" \"chunk_size\": 1024, \"chunk_overlap\": 128,\n",
" },\n",
"]\n",
"\n",
"# ── Other RAGBench datasets ────────────────────────────────────────────────────\n",
"# Add combos here for the remaining RAGBench datasets.\n",
"# First verify the HF config name:\n",
"# load_dataset(\"rungalileo/ragbench\", \"<name>\", split=\"test\")\n",
"#\n",
"# Most RAGBench datasets are short retrieved passages → use passthrough.\n",
"# Switch to recursive only if source passages exceed the model's max_seq_length.\n",
"#\n",
"# Minimum baseline per dataset: MiniLM (floor) + one stronger model.\n",
"#\n",
"# Example for FiQA:\n",
"# {\n",
"# \"dataset\": \"fiqa\", \"domain\": \"financial\",\n",
"# \"display_name\": \"FiQA (Financial)\",\n",
"# \"corpus_scope\": \"test\",\n",
"# \"embedding_model\": \"sentence-transformers/all-MiniLM-L6-v2\",\n",
"# \"embedding_display_name\": \"MiniLM L6 v2\",\n",
"# \"model_revision\": \"main\",\n",
"# \"query_prefix\": \"\", \"passage_prefix\": \"\",\n",
"# \"chunking_strategy\": \"passthrough\",\n",
"# \"chunk_size\": 0, \"chunk_overlap\": 0,\n",
"# },\n",
"OTHER_COMBOS = [] # <── extend with other RAGBench dataset combos\n",
"\n",
"COMBOS = COVIDQA_COMBOS + CUAD_COMBOS + OTHER_COMBOS\n",
"\n",
"print(f\"Repo : {HF_REPO}\")\n",
"print(f\"CovidQA : {len(COVIDQA_COMBOS)} combos\")\n",
"print(f\"CUAD : {len(CUAD_COMBOS)} combos\")\n",
"print(f\"Other : {len(OTHER_COMBOS)} combos\")\n",
"print(f\"Total : {len(COMBOS)} combos\")"
]
},
{
"cell_type": "markdown",
"id": "md-auth",
"metadata": {},
"source": [
"## Cell 3 — Authenticate with Hugging Face\n",
"\n",
"Store your token via the 🔑 Colab secrets panel (key name `HF_TOKEN`).\n",
"Never paste a token directly in a cell you might share or commit."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-auth",
"metadata": {},
"outputs": [],
"source": [
"from huggingface_hub import HfApi, login\n",
"\n",
"try:\n",
" from google.colab import userdata\n",
" HF_TOKEN = userdata.get(\"HF_TOKEN\")\n",
"except Exception:\n",
" HF_TOKEN = \"\" # <── paste token here ONLY for quick local tests, never commit\n",
"\n",
"login(token=HF_TOKEN, add_to_git_credential=False)\n",
"api = HfApi()\n",
"\n",
"api.create_repo(repo_id=HF_REPO, repo_type=\"dataset\", exist_ok=True)\n",
"print(f\"Repo ready: https://huggingface.co/datasets/{HF_REPO}\")"
]
},
{
"cell_type": "markdown",
"id": "md-helpers",
"metadata": {},
"source": [
"## Cell 4 — Helper functions\n",
"\n",
"### Cache key design\n",
"\n",
"`make_cache_key()` hashes **every field that materially changes the index**:\n",
"`schema_version`, `dataset`, `corpus_scope`, `embedding_model`, `model_revision`,\n",
"`query_prefix`, `passage_prefix`, `chunking_strategy`, `chunk_size`, `chunk_overlap`.\n",
"\n",
"Any change to any of these fields produces a new `index_id`. This means:\n",
"- Fixing a bug → increment `SCHEMA_VERSION` → all combos get new IDs automatically.\n",
"- Expanding corpus → change `corpus_scope` → new ID, old index untouched.\n",
"- Adding a prefix → new ID, no silent overwrites.\n",
"\n",
"### Corpus loading\n",
"\n",
"`load_ragbench()` accepts a `corpus_scope` string and loads the appropriate splits\n",
"from `SPLIT_MAP`. Multiple splits are concatenated before deduplication so the\n",
"knowledge base contains only unique documents.\n",
"\n",
"### Chunking dispatch\n",
"\n",
"`chunk_documents()` is a dispatcher: it reads `combo['chunking_strategy']` and calls\n",
"the matching private function. An unknown strategy raises `ValueError` immediately\n",
"so no compute is wasted. Currently implemented:\n",
"\n",
"- `passthrough` — each source document becomes exactly one chunk (no splitting).\n",
"- `recursive` — tokenizer-aware `RecursiveCharacterTextSplitter`. Uses the embedding\n",
" model's own tokenizer so `chunk_size` is in the same token space as `max_seq_length`.\n",
"\n",
"### Dense index with passage prefix\n",
"\n",
"`build_dense_index()` prepends `passage_prefix` to every chunk before encoding.\n",
"Embeddings are L2-normalised so inner-product search is equivalent to cosine similarity.\n",
"The matching `query_prefix` must be applied in the app's retriever at query time.\n",
"\n",
"### BM25 tokenisation\n",
"\n",
"`build_bm25()` uses `re.findall(r\"\\w+\", text.lower())` instead of `str.split()`.\n",
"This separates punctuation from tokens — important for legal text where terms like\n",
"`clause,` or `§3.1` would otherwise not match their clean forms.\n",
"\n",
"### Artifact schema (meta.json)\n",
"\n",
"Every uploaded artefact folder includes a `meta.json` with the full provenance:\n",
"schema version, corpus scope, model revision, effective sequence limit, both prefixes,\n",
"chunking parameters, normalisation, similarity metric, counts, and build timestamp."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-helpers",
"metadata": {},
"outputs": [],
"source": [
"import hashlib\n",
"import io\n",
"import json\n",
"import pickle\n",
"import re\n",
"import shutil\n",
"from datetime import datetime, timezone\n",
"from pathlib import Path\n",
"\n",
"import faiss\n",
"import numpy as np\n",
"import pandas as pd\n",
"from datasets import load_dataset\n",
"from rank_bm25 import BM25Okapi\n",
"from sentence_transformers import SentenceTransformer\n",
"from langchain_text_splitters import RecursiveCharacterTextSplitter\n",
"\n",
"\n",
"# ── Cache key ──────────────────────────────────────────────────────────────────\n",
"def make_cache_key(combo: dict) -> str:\n",
" \"\"\"Deterministic 12-char hash covering every field that defines index content.\"\"\"\n",
" parts = [\n",
" str(SCHEMA_VERSION),\n",
" combo[\"dataset\"],\n",
" combo[\"corpus_scope\"],\n",
" combo[\"embedding_model\"],\n",
" combo.get(\"model_revision\", \"main\"),\n",
" combo.get(\"query_prefix\", \"\"),\n",
" combo.get(\"passage_prefix\", \"\"),\n",
" combo[\"chunking_strategy\"],\n",
" str(combo[\"chunk_size\"]),\n",
" str(combo[\"chunk_overlap\"]),\n",
" ]\n",
" raw = \"|\".join(parts)\n",
" return hashlib.md5(raw.encode(\"utf-8\")).hexdigest()[:12]\n",
"\n",
"\n",
"# ── RAGBench loader ────────────────────────────────────────────────────────────\n",
"def load_ragbench(dataset_name: str, corpus_scope: str) -> pd.DataFrame:\n",
" \"\"\"Load RAGBench rows from the splits specified by corpus_scope.\"\"\"\n",
" splits = SPLIT_MAP[corpus_scope]\n",
" frames = []\n",
" for s in splits:\n",
" print(f\" Loading rungalileo/ragbench [{dataset_name}] split={s} ...\")\n",
" ds = load_dataset(\"rungalileo/ragbench\", dataset_name, split=s)\n",
" frames.append(ds.to_pandas())\n",
" df = pd.concat(frames, ignore_index=True)\n",
" print(f\" {len(df):,} rows, columns: {list(df.columns)}\")\n",
" return df\n",
"\n",
"\n",
"# ── Document deduplication ─────────────────────────────────────────────────────\n",
"def build_knowledge_base(df: pd.DataFrame) -> pd.DataFrame:\n",
" \"\"\"Flatten list-of-contexts columns into one row per unique document.\"\"\"\n",
" ctx_col = next(\n",
" (c for c in [\"documents\", \"context\", \"contexts\"] if c in df.columns), None\n",
" )\n",
" if ctx_col is None:\n",
" raise ValueError(f\"No context column found. Available: {list(df.columns)}\")\n",
" seen, rows = set(), []\n",
" for cell in df[ctx_col]:\n",
" if isinstance(cell, (list, np.ndarray)):\n",
" docs = list(cell)\n",
" elif isinstance(cell, str):\n",
" docs = [cell]\n",
" else:\n",
" docs = list(cell) if hasattr(cell, \"__iter__\") else [cell]\n",
" for doc in docs:\n",
" text = doc.strip() if isinstance(doc, str) else str(doc).strip()\n",
" if not text:\n",
" continue\n",
" h = hashlib.md5(text.encode()).hexdigest()\n",
" if h not in seen:\n",
" seen.add(h)\n",
" rows.append({\"doc_id\": h[:12], \"text\": text})\n",
" docs_df = pd.DataFrame(rows)\n",
" print(f\" {len(docs_df):,} unique documents\")\n",
" return docs_df\n",
"\n",
"\n",
"# ── Chunking strategies ────────────────────────────────────────────────────────\n",
"def _chunk_passthrough(docs_df: pd.DataFrame) -> pd.DataFrame:\n",
" \"\"\"Keep each source passage as exactly one chunk (no splitting).\"\"\"\n",
" rows = [\n",
" {\n",
" \"chunk_id\": f\"{row['doc_id']}_0\",\n",
" \"doc_id\": row[\"doc_id\"],\n",
" \"chunk_idx\": i,\n",
" \"text\": row[\"text\"],\n",
" }\n",
" for i, (_, row) in enumerate(docs_df.iterrows())\n",
" ]\n",
" return pd.DataFrame(rows).reset_index(drop=True)\n",
"\n",
"\n",
"def _chunk_recursive(\n",
" docs_df: pd.DataFrame,\n",
" model: SentenceTransformer,\n",
" chunk_size: int,\n",
" chunk_overlap: int,\n",
") -> pd.DataFrame:\n",
" \"\"\"Tokenizer-aware recursive splitting. chunk_size is in model token space.\"\"\"\n",
" splitter = RecursiveCharacterTextSplitter.from_huggingface_tokenizer(\n",
" model.tokenizer,\n",
" chunk_size=chunk_size,\n",
" chunk_overlap=chunk_overlap,\n",
" )\n",
" rows = []\n",
" for _, row in docs_df.iterrows():\n",
" for i, chunk in enumerate(splitter.split_text(row[\"text\"])):\n",
" rows.append({\n",
" \"chunk_id\": f\"{row['doc_id']}_{i}\",\n",
" \"doc_id\": row[\"doc_id\"],\n",
" \"chunk_idx\": len(rows),\n",
" \"text\": chunk,\n",
" })\n",
" return pd.DataFrame(rows).reset_index(drop=True)\n",
"\n",
"\n",
"def chunk_documents(\n",
" docs_df: pd.DataFrame,\n",
" model: SentenceTransformer,\n",
" combo: dict,\n",
") -> pd.DataFrame:\n",
" \"\"\"Dispatch to the chunking strategy named in combo['chunking_strategy'].\"\"\"\n",
" strategy = combo[\"chunking_strategy\"]\n",
" if strategy == \"passthrough\":\n",
" chunks_df = _chunk_passthrough(docs_df)\n",
" elif strategy == \"recursive\":\n",
" chunks_df = _chunk_recursive(\n",
" docs_df, model, combo[\"chunk_size\"], combo[\"chunk_overlap\"]\n",
" )\n",
" elif strategy == \"parent_child\":\n",
" raise NotImplementedError(\n",
" \"parent_child requires retrieval + generation pipeline changes. \"\n",
" \"Implement after the dense/BM25 baseline is stable.\"\n",
" )\n",
" else:\n",
" raise ValueError(\n",
" f\"Unknown chunking_strategy: {strategy!r}. \"\n",
" \"Valid options: passthrough, recursive, parent_child\"\n",
" )\n",
" print(\n",
" f\" {len(chunks_df):,} chunks from {len(docs_df):,} docs \"\n",
" f\"(strategy={strategy}, chunk_size={combo['chunk_size']}, \"\n",
" f\"overlap={combo['chunk_overlap']})\"\n",
" )\n",
" return chunks_df\n",
"\n",
"\n",
"# ── Pre-build validation ───────────────────────────────────────────────────────\n",
"def validate_combo(combo: dict, model: SentenceTransformer) -> None:\n",
" \"\"\"Raise before encoding if chunk_size would cause silent text truncation.\n",
"\n",
" SentenceTransformer.encode() silently drops tokens beyond max_seq_length.\n",
" This guard surfaces the problem before any compute is spent.\n",
" \"\"\"\n",
" if combo[\"chunking_strategy\"] == \"passthrough\":\n",
" return # no fixed chunk size; individual long passages are a data concern\n",
" if combo[\"chunk_size\"] > model.max_seq_length:\n",
" raise ValueError(\n",
" f\"chunk_size={combo['chunk_size']} exceeds \"\n",
" f\"{combo['embedding_model']} max_seq_length={model.max_seq_length}. \"\n",
" \"Reduce chunk_size or use a model with a larger limit.\"\n",
" )\n",
"\n",
"\n",
"# ── Dense index ────────────────────────────────────────────────────────────────\n",
"def build_dense_index(\n",
" chunks_df: pd.DataFrame,\n",
" model: SentenceTransformer,\n",
" passage_prefix: str = \"\",\n",
"):\n",
" \"\"\"Encode chunks with optional passage_prefix; build FAISS inner-product index.\n",
"\n",
" Embeddings are L2-normalised: inner-product == cosine similarity.\n",
" The passage_prefix is required for asymmetric models (e.g. BGE v1.5).\n",
" The matching query_prefix must be applied in the app retriever at query time.\n",
" \"\"\"\n",
" texts = [passage_prefix + t for t in chunks_df[\"text\"].tolist()]\n",
" print(f\" Encoding {len(texts):,} chunks ...\")\n",
" embeddings = model.encode(\n",
" texts,\n",
" batch_size=64,\n",
" show_progress_bar=True,\n",
" convert_to_numpy=True,\n",
" normalize_embeddings=True,\n",
" ).astype(\"float32\")\n",
" index = faiss.IndexFlatIP(embeddings.shape[1])\n",
" index.add(embeddings)\n",
" print(f\" FAISS index: {index.ntotal:,} vectors, dim={embeddings.shape[1]}\")\n",
" return embeddings, index\n",
"\n",
"\n",
"# ── BM25 index ─────────────────────────────────────────────────────────────────\n",
"def build_bm25(chunks_df: pd.DataFrame) -> BM25Okapi:\n",
" \"\"\"Build BM25 index with regex word tokenisation.\n",
"\n",
" re.findall(r'\\\\w+') separates punctuation from tokens — important for legal\n",
" text where terms like 'clause,' or '§3.1' must match their clean forms.\n",
" \"\"\"\n",
" tokenized = [re.findall(r\"\\w+\", t.lower()) for t in chunks_df[\"text\"]]\n",
" return BM25Okapi(tokenized)\n",
"\n",
"\n",
"# ── Artifact serialisation ─────────────────────────────────────────────────────\n",
"def save_artifacts(\n",
" out_dir: Path,\n",
" chunks_df: pd.DataFrame,\n",
" docs_df: pd.DataFrame,\n",
" embeddings: np.ndarray,\n",
" faiss_index,\n",
" bm25: BM25Okapi,\n",
" combo: dict,\n",
" index_id: str,\n",
" model: SentenceTransformer,\n",
") -> dict:\n",
" \"\"\"Write all artefacts to out_dir and return the meta dict.\"\"\"\n",
" out_dir.mkdir(parents=True, exist_ok=True)\n",
" faiss.write_index(faiss_index, str(out_dir / \"faiss.index\"))\n",
" np.save(str(out_dir / \"embeddings.npy\"), embeddings)\n",
" with open(out_dir / \"bm25.pkl\", \"wb\") as f:\n",
" pickle.dump(bm25, f)\n",
" chunks_df.to_parquet(out_dir / \"chunks.parquet\", index=False)\n",
" docs_df.to_parquet(out_dir / \"docs.parquet\", index=False)\n",
"\n",
" meta = {\n",
" \"schema_version\": SCHEMA_VERSION,\n",
" \"index_id\": index_id,\n",
" \"dataset_repo\": \"rungalileo/ragbench\",\n",
" \"dataset_config\": combo[\"dataset\"],\n",
" \"corpus_scope\": combo[\"corpus_scope\"],\n",
" \"evaluation_note\": (\n",
" \"Closed retrieval: documents from the selected corpus_scope only. \"\n",
" \"Metrics measure relative quality across configs, \"\n",
" \"not open-corpus absolute recall.\"\n",
" ),\n",
" \"model_name\": combo[\"embedding_model\"],\n",
" \"model_revision\": combo.get(\"model_revision\", \"main\"),\n",
" \"effective_max_seq_length\": model.max_seq_length,\n",
" \"query_prefix\": combo.get(\"query_prefix\", \"\"),\n",
" \"passage_prefix\": combo.get(\"passage_prefix\", \"\"),\n",
" \"chunking_strategy\": combo[\"chunking_strategy\"],\n",
" \"chunk_size\": combo[\"chunk_size\"],\n",
" \"chunk_overlap\": combo[\"chunk_overlap\"],\n",
" \"normalization\": \"l2\",\n",
" \"similarity\": \"inner_product\",\n",
" \"domain\": combo[\"domain\"],\n",
" \"n_chunks\": len(chunks_df),\n",
" \"n_docs\": len(docs_df),\n",
" \"build_date\": datetime.now(timezone.utc).isoformat(),\n",
" }\n",
" with open(out_dir / \"meta.json\", \"w\") as f:\n",
" json.dump(meta, f, indent=2)\n",
" print(f\" Saved to {out_dir}/\")\n",
" return meta\n",
"\n",
"\n",
"print(\"Helper functions loaded.\")"
]
},
{
"cell_type": "markdown",
"id": "md-index-config",
"metadata": {},
"source": [
"## Cell 5 — `index_config.json` helpers\n",
"\n",
"`index_config.json` is the registry file in the HF dataset repo that the Gradio app\n",
"reads to populate its dropdowns. It stores the full tree of\n",
"`dataset → embedding_model → chunk_config → index_id` mappings.\n",
"\n",
"After every successful build the notebook:\n",
"1. Downloads the current config from HF.\n",
"2. Upserts the new `chunk_config` entry (replacing any existing entry with the same\n",
" `chunk_size / chunk_overlap / chunking_strategy / corpus_scope` tuple).\n",
"3. Re-uploads the updated config.\n",
"\n",
"The `register_combo()` function now includes `corpus_scope`, `query_prefix`,\n",
"`passage_prefix`, and `build_date` in the registered entry so the app can\n",
"read them from the config without loading `meta.json`."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-index-config",
"metadata": {},
"outputs": [],
"source": [
"import copy\n",
"\n",
"INDEX_CONFIG_FILENAME = \"index_config.json\"\n",
"\n",
"_EMPTY_CONFIG = {\n",
" \"schema_version\": 1,\n",
" \"hub_repo\": HF_REPO,\n",
" \"datasets\": {},\n",
" \"runtime_options\": {\n",
" \"retrieval_strategies\": [\n",
" {\"id\": \"dense\", \"display_name\": \"Dense only\"},\n",
" {\"id\": \"bm25\", \"display_name\": \"BM25 only\"},\n",
" {\"id\": \"hybrid\", \"display_name\": \"Hybrid (Dense + BM25, RRF)\"},\n",
" ],\n",
" \"query_rewrite_strategies\": [\n",
" {\"id\": \"none\", \"display_name\": \"None (raw question)\"},\n",
" {\"id\": \"rewrite\", \"display_name\": \"Query rewrite\"},\n",
" {\"id\": \"multiquery\", \"display_name\": \"Multi-query expansion\"},\n",
" {\"id\": \"hyde\", \"display_name\": \"HyDE (dense)\"},\n",
" {\"id\": \"hyde_hybrid\", \"display_name\": \"HyDE + BM25 (hybrid)\"},\n",
" ],\n",
" \"rerankers\": [\n",
" {\"id\": \"off\", \"display_name\": \"Off\"},\n",
" {\"id\": \"cross-encoder/ms-marco-MiniLM-L-6-v2\", \"display_name\": \"MiniLM-L-6-v2 cross-encoder\"},\n",
" ],\n",
" \"llm_models\": [\n",
" {\"id\": \"llama-3.1-8b-instant\", \"display_name\": \"Llama 3.1 8B Instant\"},\n",
" {\"id\": \"llama-3.3-70b-versatile\", \"display_name\": \"Llama 3.3 70B Versatile\"},\n",
" ],\n",
" },\n",
"}\n",
"\n",
"\n",
"def download_index_config() -> dict:\n",
" \"\"\"Download index_config.json from HF repo, or return a fresh empty config.\"\"\"\n",
" try:\n",
" from huggingface_hub import hf_hub_download\n",
" path = hf_hub_download(\n",
" repo_id=HF_REPO, repo_type=\"dataset\",\n",
" filename=INDEX_CONFIG_FILENAME, token=HF_TOKEN,\n",
" )\n",
" with open(path) as f:\n",
" cfg = json.load(f)\n",
" print(\" Downloaded existing index_config.json\")\n",
" return cfg\n",
" except Exception:\n",
" print(\" index_config.json not found — starting fresh.\")\n",
" cfg = copy.deepcopy(_EMPTY_CONFIG)\n",
" cfg[\"hub_repo\"] = HF_REPO\n",
" return cfg\n",
"\n",
"\n",
"def register_combo(cfg: dict, combo: dict, meta: dict) -> dict:\n",
" \"\"\"Upsert a chunk_config entry into the in-memory config dict.\"\"\"\n",
" cfg = copy.deepcopy(cfg)\n",
" cfg[\"hub_repo\"] = HF_REPO\n",
"\n",
" ds_key = combo[\"dataset\"]\n",
" em_key = combo[\"embedding_model\"]\n",
" index_id = meta[\"index_id\"]\n",
"\n",
" ds_entry = cfg.setdefault(\"datasets\", {}).setdefault(ds_key, {\n",
" \"display_name\": combo[\"display_name\"],\n",
" \"domain\": combo[\"domain\"],\n",
" \"embedding_models\": {},\n",
" })\n",
" ds_entry[\"display_name\"] = combo[\"display_name\"]\n",
" ds_entry[\"domain\"] = combo[\"domain\"]\n",
"\n",
" em_entry = ds_entry.setdefault(\"embedding_models\", {}).setdefault(em_key, {\n",
" \"display_name\": combo[\"embedding_display_name\"],\n",
" \"chunk_configs\": [],\n",
" })\n",
" em_entry[\"display_name\"] = combo[\"embedding_display_name\"]\n",
"\n",
" new_entry = {\n",
" \"chunk_size\": combo[\"chunk_size\"],\n",
" \"chunk_overlap\": combo[\"chunk_overlap\"],\n",
" \"chunking_strategy\": combo[\"chunking_strategy\"],\n",
" \"corpus_scope\": combo[\"corpus_scope\"],\n",
" \"query_prefix\": combo.get(\"query_prefix\", \"\"),\n",
" \"passage_prefix\": combo.get(\"passage_prefix\", \"\"),\n",
" \"index_id\": index_id,\n",
" \"subfolder\": index_id,\n",
" \"files\": ARTIFACT_FILES,\n",
" \"n_chunks\": meta[\"n_chunks\"],\n",
" \"n_docs\": meta[\"n_docs\"],\n",
" \"build_date\": meta[\"build_date\"],\n",
" }\n",
"\n",
" # Replace any existing entry with the same (size, overlap, strategy, scope).\n",
" cfgs = em_entry[\"chunk_configs\"]\n",
" cfgs[:] = [\n",
" c for c in cfgs\n",
" if not (\n",
" c[\"chunk_size\"] == combo[\"chunk_size\"]\n",
" and c[\"chunk_overlap\"] == combo[\"chunk_overlap\"]\n",
" and c[\"chunking_strategy\"] == combo[\"chunking_strategy\"]\n",
" and c.get(\"corpus_scope\", \"train\") == combo[\"corpus_scope\"]\n",
" )\n",
" ]\n",
" cfgs.append(new_entry)\n",
" return cfg\n",
"\n",
"\n",
"def upload_index_config(cfg: dict) -> None:\n",
" buf = json.dumps(cfg, indent=2).encode(\"utf-8\")\n",
" api.upload_file(\n",
" path_or_fileobj=io.BytesIO(buf),\n",
" path_in_repo=INDEX_CONFIG_FILENAME,\n",
" repo_id=HF_REPO,\n",
" repo_type=\"dataset\",\n",
" token=HF_TOKEN,\n",
" commit_message=\"chore: update index_config.json\",\n",
" )\n",
" print(f\" Uploaded {INDEX_CONFIG_FILENAME} to {HF_REPO}\")\n",
"\n",
"\n",
"print(\"index_config.json helpers loaded.\")"
]
},
{
"cell_type": "markdown",
"id": "md-validate",
"metadata": {},
"source": [
"## Cell 6 — Validation: check all combos before any compute\n",
"\n",
"### Why validate first?\n",
"\n",
"`SentenceTransformer.encode()` silently truncates tokens beyond `model.max_seq_length`.\n",
"For `all-MiniLM-L6-v2` (limit = 256) with `chunk_size = 512`, the index builds without\n",
"error but encodes only the first 256 tokens of every chunk — the rest is silently dropped.\n",
"The resulting metrics look fine but the comparison across models is meaningless.\n",
"\n",
"There are three separate limits in every HuggingFace embedding model:\n",
"\n",
"| Property | Description |\n",
"|---|---|\n",
"| `config.max_position_embeddings` | Architectural capacity of the transformer |\n",
"| `tokenizer.model_max_length` | Tokenizer-level limit (may be 512 even for 256-limit models) |\n",
"| `model.max_seq_length` | **Operative limit** — what `SentenceTransformer.encode()` actually uses |\n",
"\n",
"This cell uses known `max_seq_length` values to validate all combos without loading any model.\n",
"The runtime `validate_combo()` call (Cell 9) performs the definitive check against the\n",
"actually-loaded model.\n",
"\n",
"**Run this cell before running any build. Fix any `EXCEEDS LIMIT` rows in the COMBOS config.**"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-validate",
"metadata": {},
"outputs": [],
"source": [
"# Known effective max_seq_length values (SentenceTransformer.max_seq_length).\n",
"# These are the operative limits for encode() — NOT tokenizer.model_max_length.\n",
"KNOWN_MAX_SEQ_LENGTHS = {\n",
" \"sentence-transformers/all-MiniLM-L6-v2\": 256,\n",
" \"sentence-transformers/all-mpnet-base-v2\": 384,\n",
" \"BAAI/bge-base-en-v1.5\": 512,\n",
" \"BAAI/bge-large-en-v1.5\": 512,\n",
" \"thenlper/gte-large\": 512,\n",
" \"Alibaba-NLP/gte-modernbert-base\": 8192,\n",
"}\n",
"\n",
"print(f\"{'IDX':>3} {'STATUS':<18} {'MODEL':<38} {'LIMIT':>6} {'CHUNK':>6}\")\n",
"print(\"-\" * 82)\n",
"issues = []\n",
"for i, c in enumerate(COMBOS):\n",
" strategy = c[\"chunking_strategy\"]\n",
" chunk_size = c[\"chunk_size\"]\n",
" model_id = c[\"embedding_model\"]\n",
" limit = KNOWN_MAX_SEQ_LENGTHS.get(model_id, \"?\")\n",
"\n",
" if strategy == \"passthrough\":\n",
" status = \"OK (passthrough)\"\n",
" elif limit == \"?\":\n",
" status = \"? limit unknown\"\n",
" elif chunk_size > limit:\n",
" status = \"EXCEEDS LIMIT\"\n",
" issues.append(i)\n",
" else:\n",
" status = \"valid\"\n",
"\n",
" short = model_id.split(\"/\")[-1]\n",
" print(f\"{i:>3} {status:<18} {short:<38} {str(limit):>6} {chunk_size:>6}\")\n",
"\n",
"print()\n",
"if issues:\n",
" print(f\"FAIL: {len(issues)} combo(s) have invalid chunk sizes: indices {issues}\")\n",
" print(\"Fix chunk_size in COMBOS (Cell 2) before running any build.\")\n",
"else:\n",
" print(\"All combos valid. Safe to build.\")"
]
},
{
"cell_type": "markdown",
"id": "md-status",
"metadata": {},
"source": [
"## Cell 7 — Status: which combos are done / pending\n",
"\n",
"Queries the HF repo for already-uploaded subfolders and matches them against\n",
"the current COMBOS list. Run this at the start of every session to choose\n",
"which `COMBO_IDX` to build next."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-status",
"metadata": {},
"outputs": [],
"source": [
"from huggingface_hub import list_repo_files\n",
"\n",
"\n",
"def _uploaded_index_ids() -> set:\n",
" \"\"\"Return subfolder names already present in the HF dataset repo.\"\"\"\n",
" try:\n",
" files = list(list_repo_files(repo_id=HF_REPO, repo_type=\"dataset\", token=HF_TOKEN))\n",
" return {f.split(\"/\")[0] for f in files if \"/\" in f}\n",
" except Exception:\n",
" return set()\n",
"\n",
"\n",
"done_ids = _uploaded_index_ids()\n",
"\n",
"print(f\"{'IDX':>3} {'STATUS':<10} {'INDEX_ID':<14} COMBO\")\n",
"print(\"-\" * 80)\n",
"for i, c in enumerate(COMBOS):\n",
" idx_id = make_cache_key(c)\n",
" status = \"done\" if idx_id in done_ids else \"pending\"\n",
" short = c[\"embedding_model\"].split(\"/\")[-1]\n",
" label = (\n",
" f\"{c['dataset']} / {short} / \"\n",
" f\"{c['chunking_strategy']}:{c['chunk_size']}-{c['chunk_overlap']} / \"\n",
" f\"scope={c['corpus_scope']}\"\n",
" )\n",
" print(f\"{i:>3} {status:<10} {idx_id:<14} {label}\")\n",
"\n",
"n_done = sum(1 for c in COMBOS if make_cache_key(c) in done_ids)\n",
"n_pending = len(COMBOS) - n_done\n",
"print(f\"\\n{n_done} done / {n_pending} pending / {len(COMBOS)} total\")"
]
},
{
"cell_type": "markdown",
"id": "md-pick",
"metadata": {},
"source": [
"## Cell 8 — Pick the combo to build\n",
"\n",
"Set `COMBO_IDX` to the row number shown in Cell 7 that you want to build,\n",
"then run **Cells 8–12** in sequence."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-pick",
"metadata": {},
"outputs": [],
"source": [
"COMBO_IDX = 0 # <── CHANGE THIS\n",
"\n",
"combo = COMBOS[COMBO_IDX]\n",
"INDEX_ID = make_cache_key(combo)\n",
"\n",
"print(f\"Selected combo {COMBO_IDX}:\")\n",
"print(f\" dataset : {combo['dataset']} ({combo['domain']})\")\n",
"print(f\" corpus_scope : {combo['corpus_scope']}\")\n",
"print(f\" embedding_model : {combo['embedding_model']}\")\n",
"print(f\" model_revision : {combo.get('model_revision', 'main')}\")\n",
"print(f\" query_prefix : {combo.get('query_prefix', '')!r}\")\n",
"print(f\" passage_prefix : {combo.get('passage_prefix', '')!r}\")\n",
"print(f\" chunking_strategy : {combo['chunking_strategy']}\")\n",
"print(f\" chunk_size : {combo['chunk_size']}\")\n",
"print(f\" chunk_overlap : {combo['chunk_overlap']}\")\n",
"print(f\" index_id (key) : {INDEX_ID}\")\n",
"\n",
"if INDEX_ID in done_ids:\n",
" print(\"\\nWARNING: This combo is already uploaded.\")\n",
" print(\"Continue to re-build and overwrite, or change COMBO_IDX.\")"
]
},
{
"cell_type": "markdown",
"id": "md-build-kb",
"metadata": {},
"source": [
"## Cell 9 — Load dataset, build knowledge base, chunk\n",
"\n",
"Steps:\n",
"1. Load RAGBench rows from the splits defined by `corpus_scope`.\n",
"2. Flatten the per-row context lists into a deduplicated document table.\n",
"3. Load the embedding model (needed for the tokenizer in recursive chunking).\n",
"4. Run the pre-build validation guard — raises immediately if `chunk_size` exceeds\n",
" the model's effective sequence limit.\n",
"5. Chunk documents using the strategy declared in the combo."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-build-kb",
"metadata": {},
"outputs": [],
"source": [
"print(f\"Loading {combo['dataset']} [{combo['corpus_scope']}] ...\")\n",
"full_df = load_ragbench(combo[\"dataset\"], combo[\"corpus_scope\"])\n",
"\n",
"print(\"\\nDeduplicating documents ...\")\n",
"docs_df = build_knowledge_base(full_df)\n",
"\n",
"print(f\"\\nLoading embedding model: {combo['embedding_model']} ...\")\n",
"embed_model = SentenceTransformer(\n",
" combo[\"embedding_model\"],\n",
" revision=combo.get(\"model_revision\", \"main\"),\n",
")\n",
"print(f\" max_seq_length : {embed_model.max_seq_length}\")\n",
"print(f\" embedding_dim : {embed_model.get_sentence_embedding_dimension()}\")\n",
"\n",
"print(\"\\nValidating combo against model limits ...\")\n",
"validate_combo(combo, embed_model)\n",
"print(\" Validation passed\")\n",
"\n",
"print(\"\\nChunking documents ...\")\n",
"chunks_df = chunk_documents(docs_df, embed_model, combo)"
]
},
{
"cell_type": "markdown",
"id": "md-build-idx",
"metadata": {},
"source": [
"## Cell 10 — Build dense and BM25 indices\n",
"\n",
"**Dense index:** each chunk is encoded with the embedding model. If the combo\n",
"specifies a `passage_prefix`, it is prepended to every chunk before encoding.\n",
"Vectors are L2-normalised so FAISS inner-product search is equivalent to cosine\n",
"similarity. `IndexFlatIP` performs exact search — no approximation.\n",
"\n",
"**BM25 index:** built with improved regex tokenisation. BM25 has no incremental\n",
"update API; if the corpus scope is expanded later the BM25 index must be rebuilt\n",
"from scratch (which is why the full pipeline is re-run for each combo)."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-build-idx",
"metadata": {},
"outputs": [],
"source": [
"print(\"Building dense index ...\")\n",
"embeddings, faiss_index = build_dense_index(\n",
" chunks_df,\n",
" embed_model,\n",
" passage_prefix=combo.get(\"passage_prefix\", \"\"),\n",
")\n",
"\n",
"print(\"\\nBuilding BM25 index ...\")\n",
"bm25 = build_bm25(chunks_df)\n",
"print(f\" BM25 done ({len(chunks_df):,} tokenized chunks)\")"
]
},
{
"cell_type": "markdown",
"id": "md-save",
"metadata": {},
"source": [
"## Cell 11 — Save artifacts locally and upload to HF Hub\n",
"\n",
"Artifacts saved per combo:\n",
"\n",
"| File | Contents |\n",
"|---|---|\n",
"| `faiss.index` | FAISS `IndexFlatIP` — serialised binary |\n",
"| `embeddings.npy` | Float32 numpy array of L2-normalised vectors |\n",
"| `bm25.pkl` | Pickled `BM25Okapi` object |\n",
"| `chunks.parquet` | Chunk table: `chunk_id`, `doc_id`, `chunk_idx`, `text` |\n",
"| `docs.parquet` | Document table: `doc_id`, `text` |\n",
"| `meta.json` | Full provenance record (see Cell 4 for schema) |\n",
"\n",
"All six files are uploaded to `{HF_REPO}/{index_id}/` in one `upload_folder` call."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-save",
"metadata": {},
"outputs": [],
"source": [
"OUT_DIR = Path(f\"/tmp/rag_indices/{INDEX_ID}\")\n",
"\n",
"print(f\"Saving artifacts to {OUT_DIR} ...\")\n",
"meta = save_artifacts(\n",
" OUT_DIR, chunks_df, docs_df,\n",
" embeddings, faiss_index, bm25,\n",
" combo, INDEX_ID, embed_model,\n",
")\n",
"\n",
"print(f\"\\nUploading {OUT_DIR} → {HF_REPO}/{INDEX_ID} ...\")\n",
"api.upload_folder(\n",
" folder_path=str(OUT_DIR),\n",
" path_in_repo=INDEX_ID,\n",
" repo_id=HF_REPO,\n",
" repo_type=\"dataset\",\n",
" token=HF_TOKEN,\n",
" commit_message=(\n",
" f\"add {INDEX_ID} \"\n",
" f\"({combo['dataset']} / {combo['embedding_model']} / \"\n",
" f\"{combo['chunking_strategy']}:{combo['chunk_size']}-{combo['chunk_overlap']} / \"\n",
" f\"scope={combo['corpus_scope']})\"\n",
" ),\n",
")\n",
"print(\" Upload complete.\")"
]
},
{
"cell_type": "markdown",
"id": "md-update-config",
"metadata": {},
"source": [
"## Cell 12 — Update `index_config.json`\n",
"\n",
"Downloads the current registry, upserts the new combo entry, and re-uploads.\n",
"The Gradio app reads this file to build its dataset / model / chunk dropdowns."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-update-config",
"metadata": {},
"outputs": [],
"source": [
"print(\"Updating index_config.json ...\")\n",
"current_cfg = download_index_config()\n",
"updated_cfg = register_combo(current_cfg, combo, meta)\n",
"upload_index_config(updated_cfg)\n",
"\n",
"# Refresh done_ids so Cell 7 is accurate if re-run in the same session.\n",
"done_ids = _uploaded_index_ids()\n",
"\n",
"print(\"\\n── Registered chunk_config entry ──\")\n",
"ds_cfg = updated_cfg[\"datasets\"].get(combo[\"dataset\"], {})\n",
"em_cfg = ds_cfg.get(\"embedding_models\", {}).get(combo[\"embedding_model\"], {})\n",
"for cc in em_cfg.get(\"chunk_configs\", []):\n",
" if cc[\"index_id\"] == INDEX_ID:\n",
" print(json.dumps(cc, indent=2))\n",
" break\n",
"\n",
"print(f\"\\nCombo {COMBO_IDX} complete. Run Cell 7 to see remaining combos.\")"
]
},
{
"cell_type": "markdown",
"id": "md-batch",
"metadata": {},
"source": [
"## Cell 13 — Batch: build all pending combos automatically\n",
"\n",
"Alternative to running Cells 8–12 one combo at a time. Discovers every pending\n",
"combo from the HF repo, then builds and uploads them in sequence.\n",
"\n",
"**Optimisation:** combos are sorted by `(dataset, corpus_scope, embedding_model)` so\n",
"each dataset is loaded once and each model is loaded once across all its chunk-size\n",
"variants. This avoids redundant downloads on free-tier Colab.\n",
"\n",
"Set `DRY_RUN = True` to preview what would be built without running any compute.\n",
"\n",
"Each combo's result is uploaded immediately on completion so progress is never lost\n",
"if the session is interrupted."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-batch",
"metadata": {},
"outputs": [],
"source": [
"import gc\n",
"from itertools import groupby\n",
"\n",
"DRY_RUN = False # set True to preview without building\n",
"\n",
"# ── Gather pending combos ──────────────────────────────────────────────────────\n",
"done_ids = _uploaded_index_ids()\n",
"pending = [\n",
" (i, make_cache_key(c), c)\n",
" for i, c in enumerate(COMBOS)\n",
" if make_cache_key(c) not in done_ids\n",
"]\n",
"\n",
"print(f\"{len(pending)} pending / {len(COMBOS)} total\")\n",
"for i, idx_id, c in pending:\n",
" short = c[\"embedding_model\"].split(\"/\")[-1]\n",
" print(\n",
" f\" [{i}] {idx_id} {c['dataset']} / {short} / \"\n",
" f\"{c['chunking_strategy']}:{c['chunk_size']}-{c['chunk_overlap']} / \"\n",
" f\"scope={c['corpus_scope']}\"\n",
" )\n",
"\n",
"if DRY_RUN:\n",
" print(\"\\nDRY_RUN=True — nothing built.\")\n",
"else:\n",
" # Sort so dataset+scope and model are each loaded once.\n",
" keyfn = lambda t: (t[2][\"dataset\"], t[2][\"corpus_scope\"], t[2][\"embedding_model\"])\n",
" pending_sorted = sorted(pending, key=keyfn)\n",
"\n",
" _ds_cache = {} # \"dataset::scope\" -> docs_df\n",
" _model_key = None\n",
" _model = None\n",
"\n",
" for (ds_name, scope, em_name), group in groupby(pending_sorted, key=keyfn):\n",
" group = list(group)\n",
" ds_scope_key = f\"{ds_name}::{scope}\"\n",
"\n",
" if ds_scope_key not in _ds_cache:\n",
" print(f\"\\n── Loading dataset: {ds_name} [{scope}] ──\")\n",
" raw_df = load_ragbench(ds_name, scope)\n",
" _ds_cache[ds_scope_key] = build_knowledge_base(raw_df)\n",
"\n",
" if _model_key != em_name:\n",
" print(f\"\\n── Loading model: {em_name} ──\")\n",
" _model = SentenceTransformer(em_name)\n",
" _model_key = em_name\n",
"\n",
" docs_df = _ds_cache[ds_scope_key]\n",
" embed_model = _model\n",
"\n",
" for combo_idx, index_id, combo in group:\n",
" print(f\"\\n{'='*60}\")\n",
" print(f\"Building [{combo_idx}] {index_id}\")\n",
" print(f\" {combo['chunking_strategy']}:{combo['chunk_size']}-{combo['chunk_overlap']}\")\n",
" print(f\"{'='*60}\")\n",
" try:\n",
" validate_combo(combo, embed_model)\n",
" chunks_df = chunk_documents(docs_df, embed_model, combo)\n",
" embeddings, faiss_index = build_dense_index(\n",
" chunks_df, embed_model,\n",
" passage_prefix=combo.get(\"passage_prefix\", \"\"),\n",
" )\n",
" bm25 = build_bm25(chunks_df)\n",
" out_dir = Path(f\"/tmp/rag_indices/{index_id}\")\n",
" meta = save_artifacts(\n",
" out_dir, chunks_df, docs_df,\n",
" embeddings, faiss_index, bm25,\n",
" combo, index_id, embed_model,\n",
" )\n",
" api.upload_folder(\n",
" folder_path=str(out_dir),\n",
" path_in_repo=index_id,\n",
" repo_id=HF_REPO,\n",
" repo_type=\"dataset\",\n",
" token=HF_TOKEN,\n",
" commit_message=(\n",
" f\"add {index_id} \"\n",
" f\"({combo['dataset']}/{combo['embedding_model']}/\"\n",
" f\"{combo['chunking_strategy']}:{combo['chunk_size']}-{combo['chunk_overlap']})\"\n",
" ),\n",
" )\n",
" current_cfg = download_index_config()\n",
" updated_cfg = register_combo(current_cfg, combo, meta)\n",
" upload_index_config(updated_cfg)\n",
" done_ids = _uploaded_index_ids()\n",
" del embeddings, faiss_index, bm25, chunks_df\n",
" shutil.rmtree(out_dir, ignore_errors=True)\n",
" gc.collect()\n",
" print(f\"[{combo_idx}] done\")\n",
" except Exception as exc:\n",
" print(f\"[{combo_idx}] FAILED: {exc}\")\n",
" print(\" Skipping to next combo ...\")\n",
"\n",
" print(\"\\nAll pending combos processed.\")"
]
},
{
"cell_type": "markdown",
"id": "md-final",
"metadata": {},
"source": [
"## Cell 14 — Final status\n",
"\n",
"Re-fetches the uploaded subfolder list from HF and shows the full progress table.\n",
"Run after each session to confirm what has been uploaded."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "cd-final",
"metadata": {},
"outputs": [],
"source": [
"done_ids = _uploaded_index_ids()\n",
"n_done = sum(1 for c in COMBOS if make_cache_key(c) in done_ids)\n",
"\n",
"print(f\"Progress: {n_done}/{len(COMBOS)} combos uploaded\\n\")\n",
"print(f\"{'IDX':>3} {'STATUS':<10} {'INDEX_ID':<14} COMBO\")\n",
"print(\"-\" * 80)\n",
"for i, c in enumerate(COMBOS):\n",
" idx_id = make_cache_key(c)\n",
" status = \"done\" if idx_id in done_ids else \"pending\"\n",
" short = c[\"embedding_model\"].split(\"/\")[-1]\n",
" label = (\n",
" f\"{c['dataset']} / {short} / \"\n",
" f\"{c['chunking_strategy']}:{c['chunk_size']}-{c['chunk_overlap']} / \"\n",
" f\"scope={c['corpus_scope']}\"\n",
" )\n",
" print(f\"{i:>3} {status:<10} {idx_id:<14} {label}\")"
]
}
]
}
|