Spaces:
Running
Running
File size: 81,320 Bytes
5ea3240 ef2c57c 5ea3240 ef2c57c 5ea3240 f0307a2 27716f7 03bed0b 27716f7 bb5d2bb 5ea3240 f79a242 b0af996 f0307a2 5ea3240 ef2c57c f4b92b8 ef2c57c 5ea3240 13784a6 f4b92b8 b0af996 f4b92b8 f0307a2 27716f7 f4b92b8 bb5d2bb f0307a2 27716f7 f4b92b8 ef2c57c f79a242 ef2c57c f4b92b8 f79a242 f0307a2 ef2c57c f0307a2 b0af996 f0307a2 b0af996 f0307a2 b0af996 ef2c57c f4b92b8 b0af996 f4b92b8 ef2c57c b0af996 f4b92b8 b0af996 ef2c57c f4b92b8 ef2c57c 13784a6 ef2c57c f79a242 3f1f29b f0307a2 f79a242 f0307a2 3f1f29b f0307a2 f79a242 f0307a2 f79a242 3f1f29b f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 ef2c57c 13784a6 ef2c57c bb5d2bb ef2c57c bb5d2bb 13784a6 ef2c57c 13784a6 ef2c57c 13784a6 ef2c57c 13784a6 ef2c57c 3f1f29b ef2c57c 13784a6 bb5d2bb 13784a6 ef2c57c f4b92b8 bb5d2bb f4b92b8 bb5d2bb f4b92b8 bb5d2bb f4b92b8 ef2c57c bb5d2bb ef2c57c bb5d2bb ef2c57c bb5d2bb ef2c57c f4b92b8 bb5d2bb ef2c57c f4b92b8 ef2c57c bb5d2bb ef2c57c 27716f7 f0307a2 ef2c57c bb5d2bb f4b92b8 ef2c57c 13784a6 ef2c57c bb5d2bb ef2c57c bb5d2bb ef2c57c 13784a6 ef2c57c 13784a6 ef2c57c bb5d2bb ef2c57c f4b92b8 bb5d2bb 13784a6 ef2c57c f0307a2 ef2c57c f0307a2 ef2c57c f0307a2 ef2c57c f0307a2 ef2c57c bb5d2bb ef2c57c 27716f7 f0307a2 ef2c57c bb5d2bb f4b92b8 ef2c57c bb5d2bb ef2c57c bb5d2bb ef2c57c bb5d2bb ef2c57c bb5d2bb 27716f7 bb5d2bb 27716f7 bb5d2bb ef2c57c 13784a6 ef2c57c bb5d2bb 27716f7 bb5d2bb ef2c57c bb5d2bb ef2c57c f4b92b8 ef2c57c 13784a6 ef2c57c 13784a6 ef2c57c 5ea3240 ef2c57c f4b92b8 ef2c57c f4b92b8 b0af996 f4b92b8 b0af996 8cfc5e2 b0af996 f79a242 b0af996 8cfc5e2 b0af996 03bed0b f4b92b8 b0af996 03bed0b f79a242 f0307a2 f4b92b8 bb5d2bb f0307a2 3f1f29b f0307a2 03bed0b f0307a2 03bed0b f0307a2 03bed0b f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f79a242 f0307a2 f4b92b8 bb5d2bb f4b92b8 27716f7 f4b92b8 b0af996 f4b92b8 f79a242 f4b92b8 03bed0b f4b92b8 27716f7 b0af996 03bed0b f79a242 f0307a2 27716f7 ef2c57c bb5d2bb 27716f7 b0af996 ef2c57c f4b92b8 ef2c57c f79a242 f4b92b8 ef2c57c f4b92b8 ef2c57c b0af996 27716f7 ef2c57c bb5d2bb ef2c57c b0af996 ef2c57c bb5d2bb ef2c57c bb5d2bb b0af996 bb5d2bb b0af996 f0307a2 ef2c57c f79a242 f0307a2 ef2c57c f4b92b8 ef2c57c b0af996 ef2c57c bb5d2bb ef2c57c b0af996 ef2c57c b0af996 ef2c57c b0af996 ef2c57c b0af996 ef2c57c b0af996 ef2c57c f4b92b8 ef2c57c f4b92b8 ef2c57c bb5d2bb f4b92b8 b0af996 f4b92b8 f0307a2 3f1f29b f0307a2 ef2c57c f4b92b8 ef2c57c f4b92b8 ef2c57c f4b92b8 f0307a2 ef2c57c b0af996 ef2c57c f4b92b8 ef2c57c f79a242 f0307a2 f4b92b8 bb5d2bb f0307a2 ef2c57c b0af996 03bed0b 3f1f29b 03bed0b f0307a2 03bed0b ef2c57c f4b92b8 13784a6 ef2c57c f79a242 f0307a2 b0af996 03bed0b b0af996 ef2c57c f4b92b8 f0307a2 ef2c57c bb5d2bb 27716f7 bb5d2bb 27716f7 f4b92b8 ef2c57c b0af996 3f1f29b f0307a2 ef2c57c 5ea3240 | 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 | from __future__ import annotations
import json
import statistics
import time
import uuid
from pathlib import Path
from typing import Any, Callable
from .context_budget import adaptive_context_budget, adaptive_retrieval_top_k, focused_context_budget
from .evidence_compression import focused_evidence_compression
from .eval_metrics import (
answer_key_match,
citation_metrics,
mean,
missing_answer_match,
percentile,
safe_div,
scalar_value_match,
source_metrics,
)
from .llm import GeminiGateway, RequestPacer
from .pipeline import RAGEngine
from .schemas import PipelineConfig, QueryPlan
from .security import prompt_injection_score
from .stress_eval import scale_stress_retrieval_eval
from .workspace import Workspace
ROOT = Path(__file__).resolve().parents[2]
BENCHMARK_PATH = ROOT / "evals" / "demo_benchmark.json"
def _load_benchmark() -> dict[str, Any]:
return json.loads(BENCHMARK_PATH.read_text(encoding="utf-8"))
def demo_benchmark_metadata() -> dict[str, Any]:
benchmark = _load_benchmark()
return {
"version": benchmark.get("version"),
"description": benchmark.get("description", ""),
"cases": {
"focused_qa": len(benchmark.get("qa_cases", [])),
"semantic_planner": len(benchmark.get("planner_cases", [])),
"corpus_overview": len(benchmark.get("overview_cases", [])),
"text2sql": len(benchmark.get("sql_cases", [])),
"hard_mode": len(benchmark.get("hard_mode_cases", [])),
"lifecycle_abstention": 2,
},
"levels": {
"Quick": "Small deployment smoke test",
"Standard": "Full deterministic benchmark, hard-mode robustness, retrieval/context/compression ablations, synthetic scale stress and release-readiness checks",
"Deep": "Calibrated Gemini judge layered onto Standard; reuses a compatible saved Standard baseline when available",
},
"default_target_rpm": 12,
"deep_judge_cases": sum(
1 for case in benchmark.get("qa_cases", []) + benchmark.get("overview_cases", []) if case.get("deep_judge")
),
"zero_gemini_ablations": ["reranker", "adaptive-context-budget", "evidence-compression", "scale-stress"],
"cache_policy": (
"RAG response cache is bypassed during fresh benchmark execution. Completed Quick/Standard/Deep reports "
"can be saved per workspace, and Deep can reuse a compatible Standard deterministic baseline."
),
}
def _document_sources(result_sources: list[dict[str, Any]], k: int = 5) -> list[str]:
return [
str(source.get("title", ""))
for source in result_sources
if source.get("type") == "document"
][:k]
def _evidence_text(result_sources: list[dict[str, Any]]) -> str:
blocks = []
for source in result_sources:
sid = source.get("id", "?")
title = source.get("title", "Source")
snippet = source.get("snippet", "")
url = source.get("url")
blocks.append(f"[{sid}] {title}\n{('URL: ' + url + chr(10)) if url else ''}{snippet}")
return "\n\n".join(blocks)
def _trace_efficiency(trace: dict[str, Any]) -> dict[str, Any]:
metrics = trace.get("metrics", {})
nodes = trace.get("nodes", [])
names = [str(node.get("node", "")) for node in nodes]
retrieve = next((node for node in reversed(nodes) if node.get("node") == "retrieve"), {})
generate = next((node for node in reversed(nodes) if node.get("node") == "generate"), {})
return {
"node_count": int(metrics.get("node_count", len(nodes)) or 0),
"llm_calls_estimate": int(metrics.get("llm_calls_estimate", 0) or 0),
"web_used": bool(metrics.get("web_used", "web" in names)),
"correction_used": bool(metrics.get("correction_used", "correct" in names)),
"abstained": bool(metrics.get("abstained", "abstain" in names)),
"cache_hit": bool(trace.get("cache_hit", False)),
"context_pruning_used": bool(retrieve.get("context_pruning_used", False)),
"context_chunks_before": int(retrieve.get("context_chunks_before", 0) or 0),
"context_chunks_after": int(retrieve.get("context_chunks_after", 0) or 0),
"context_tokens_est_before": int(retrieve.get("context_tokens_est_before", 0) or 0),
"context_tokens_est_after": int(retrieve.get("context_tokens_est_after", 0) or 0),
"context_reduction_pct": float(retrieve.get("context_reduction_pct", 0.0) or 0.0),
"manifest_included": bool(generate.get("manifest_included", False)),
"generation_prompt_tokens_est": int(generate.get("generation_prompt_tokens_est", 0) or 0),
"generation_output_tokens_est": int(generate.get("generation_output_tokens_est", 0) or 0),
"generation_total_tokens_est": int(generate.get("generation_total_tokens_est", 0) or 0),
"evidence_source_utilization_rate": float(generate.get("evidence_source_utilization_rate", 0.0) or 0.0),
"context_budget_target_chunks": int(retrieve.get("context_budget_target_chunks", 0) or 0),
"context_budget_policy": str(retrieve.get("context_budget_policy", "")),
"corpus_scale": str(retrieve.get("corpus_scale", "")),
"retrieval_top_k": int(retrieve.get("retrieval_top_k", 0) or 0),
"retrieval_confidence": float(retrieve.get("retrieval_confidence", 0.0) or 0.0),
"retrieval_score_gap": float(retrieve.get("retrieval_score_gap", 0.0) or 0.0),
"evidence_compression_used": bool(retrieve.get("evidence_compression_used", False)),
"evidence_compression_reduction_pct": float(retrieve.get("evidence_compression_reduction_pct", 0.0) or 0.0),
"evidence_tokens_est_after_compression": int(retrieve.get("evidence_tokens_est_after_compression", 0) or 0),
}
def _trace_node_times(trace: dict[str, Any], pacing_wait_ms: float = 0.0) -> dict[str, float]:
"""Return approximate service-node time with deliberate eval pacing removed.
RequestPacer sleeps happen inside the LLM node that is about to issue a
provider request. The trace records node wall time, so evaluation subtracts
the query-level deliberate pacing proportionally across nodes according to
their recorded LLM-call counts. Raw trace JSON remains unchanged.
"""
nodes = list(trace.get("nodes", []))
total_calls = sum(int(node.get("llm_calls", 0) or 0) for node in nodes)
wait_per_call = (max(0.0, float(pacing_wait_ms)) / total_calls) if total_calls else 0.0
out: dict[str, float] = {}
for node in nodes:
name = str(node.get("node", ""))
if not name:
continue
raw = float(node.get("ms", 0.0) or 0.0)
node_wait = wait_per_call * int(node.get("llm_calls", 0) or 0)
service = max(0.0, raw - node_wait)
out[name] = out.get(name, 0.0) + service
return out
def _chunk_rank_metrics(hits: list[Any], case: dict[str, Any]) -> tuple[float | None, float | None]:
terms = [str(x).lower() for x in case.get("chunk_must_contain", [])]
if not terms:
return None, None
first_rank = 0
for rank, hit in enumerate(hits[:5], start=1):
text = hit.chunk.text.lower()
if all(term in text for term in terms):
first_rank = rank
break
return (float(first_rank == 1), (1.0 / first_rank if first_rank else 0.0))
def _node_latency_summary(rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
buckets: dict[str, list[float]] = {}
for row in rows:
for node, ms in (row.get("_node_times") or {}).items():
buckets.setdefault(node, []).append(float(ms))
out = []
for node, values in buckets.items():
out.append({
"node": node,
"mean_ms": round(mean(values), 1),
"p50_ms": round(percentile(values, 0.50), 1),
"p95_ms": round(percentile(values, 0.95), 1),
"samples": len(values),
})
return sorted(out, key=lambda row: float(row["mean_ms"]), reverse=True)
def _retrieval_ablation(workspace: Workspace, qa_cases: list[dict[str, Any]]) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
for use_reranker in (False, True):
metric_buckets: dict[str, list[float]] = {
"source_recall@5": [],
"source_hit@1": [],
"source_mrr": [],
"source_ap@5": [],
"source_ndcg@5": [],
"source_duplicate_rate@5": [],
"chunk_hit@1": [],
"chunk_mrr": [],
}
latencies: list[float] = []
for case in qa_cases:
started = time.perf_counter()
hits = workspace.retriever.search(case["question"], top_k=5, use_reranker=use_reranker)
latency = (time.perf_counter() - started) * 1000
returned = [hit.chunk.source for hit in hits]
metrics = source_metrics(returned, case.get("relevant_sources", []))
chunk_hit, chunk_mrr = _chunk_rank_metrics(hits, case)
if chunk_hit is not None:
metrics["chunk_hit@1"] = chunk_hit
metrics["chunk_mrr"] = chunk_mrr
for key in metric_buckets:
if key in metrics:
metric_buckets[key].append(float(metrics[key]))
latencies.append(latency)
row = {
"configuration": "Hybrid + reranker" if use_reranker else "Hybrid RRF",
**{key: round(mean(values), 3) for key, values in metric_buckets.items()},
"median_retrieval_ms": round(statistics.median(latencies), 1) if latencies else 0.0,
}
rows.append(row)
return rows
def _context_budget_ablation(workspace: Workspace, qa_cases: list[dict[str, Any]]) -> list[dict[str, Any]]:
"""Zero-Gemini comparison of full, fixed, and adaptive context budgets."""
labels = ("Full top-k", "Fixed 3-chunk budget", "Adaptive budget")
buckets: dict[str, list[dict[str, float]]] = {label: [] for label in labels}
for case in qa_cases:
plan = QueryPlan(
route="documents",
knowledge_scope="corpus",
task_type="fact_lookup",
retrieval_strategy="semantic",
web_relevance="irrelevant",
rewritten_query=case["question"],
document_queries=[case["question"]],
)
cfg = PipelineConfig(profile="Balanced", top_k=6, use_context_pruning=True, use_adaptive_top_k=True)
effective_k = adaptive_retrieval_top_k(
cfg, plan, corpus_chunks=len(workspace.chunks), corpus_sources=len(workspace.source_profiles)
)
hits = workspace.retriever.search(case["question"], top_k=effective_k, use_reranker=False)
fixed = focused_context_budget(hits, plan, cfg)
adaptive = adaptive_context_budget(
hits, plan, cfg, corpus_chunks=len(workspace.chunks), corpus_sources=len(workspace.source_profiles)
)
variants = {
"Full top-k": (hits, len(hits), "full"),
"Fixed 3-chunk budget": (fixed.hits, fixed.target_chunks, fixed.reason),
"Adaptive budget": (adaptive.hits, adaptive.target_chunks, adaptive.reason),
}
full_chars = max(1, sum(len(hit.chunk.text or "") + len(hit.chunk.source or "") + 24 for hit in hits))
for label, (variant, target, reason) in variants.items():
returned = [hit.chunk.source for hit in variant]
metrics = source_metrics(returned, case.get("relevant_sources", []))
chars = sum(len(hit.chunk.text or "") + len(hit.chunk.source or "") + 24 for hit in variant)
buckets[label].append({
"source_precision@5": float(metrics["source_precision@5"]),
"source_recall@5": float(metrics["source_recall@5"]),
"source_hit@1": float(metrics["source_hit@1"]),
"source_mrr": float(metrics["source_mrr"]),
"context_chunks": float(len(variant)),
"target_chunks": float(target),
"context_sources": float(len(set(returned))),
"context_chars": float(chars),
"context_tokens_est": float((chars + 3) // 4),
"context_reduction_pct": float(max(0.0, 1.0 - chars / full_chars) * 100.0),
"adaptive_used": float(label == "Adaptive budget" and len(variant) < len(hits)),
})
rows: list[dict[str, Any]] = []
for label in labels:
values = buckets[label]
rows.append({
"configuration": label,
"source_precision@5": round(mean([v["source_precision@5"] for v in values]), 3),
"source_recall@5": round(mean([v["source_recall@5"] for v in values]), 3),
"source_hit@1": round(mean([v["source_hit@1"] for v in values]), 3),
"source_mrr": round(mean([v["source_mrr"] for v in values]), 3),
"median_context_chunks": round(statistics.median([v["context_chunks"] for v in values]), 1) if values else 0.0,
"median_target_chunks": round(statistics.median([v["target_chunks"] for v in values]), 1) if values else 0.0,
"median_context_sources": round(statistics.median([v["context_sources"] for v in values]), 1) if values else 0.0,
"median_context_chars": round(statistics.median([v["context_chars"] for v in values]), 1) if values else 0.0,
"median_context_tokens_est": round(statistics.median([v["context_tokens_est"] for v in values]), 1) if values else 0.0,
"median_context_reduction_pct": round(statistics.median([v["context_reduction_pct"] for v in values]), 1) if values else 0.0,
})
return rows
def _compression_signal(case: dict[str, Any], text: str) -> bool:
lower = (text or "").lower()
required = [str(term).lower() for term in case.get("chunk_must_contain", [])]
if required:
return all(term in lower for term in required)
expected = [str(term).lower() for term in case.get("expected_any", [])]
return any(term in lower for term in expected) if expected else True
def _evidence_compression_ablation(workspace: Workspace, qa_cases: list[dict[str, Any]]) -> list[dict[str, Any]]:
"""Zero-Gemini test that focused sentence compression retains labeled answer signals."""
full_tokens: list[float] = []
compressed_tokens: list[float] = []
retention: list[float] = []
reductions: list[float] = []
for case in qa_cases:
plan = QueryPlan(
route="documents", knowledge_scope="corpus", task_type="fact_lookup", retrieval_strategy="semantic",
web_relevance="irrelevant", rewritten_query=case["question"], document_queries=[case["question"]],
)
cfg = PipelineConfig(profile="Balanced", top_k=6, use_context_pruning=True, use_adaptive_top_k=True)
effective_k = adaptive_retrieval_top_k(
cfg, plan, corpus_chunks=len(workspace.chunks), corpus_sources=len(workspace.source_profiles)
)
hits = workspace.retriever.search(case["question"], top_k=effective_k, use_reranker=False)
budget = adaptive_context_budget(
hits, plan, cfg, corpus_chunks=len(workspace.chunks), corpus_sources=len(workspace.source_profiles)
)
compression = focused_evidence_compression(budget.hits, plan, query=case["question"], enabled=True)
before_text = "\n".join(hit.chunk.text or "" for hit in budget.hits)
after_text = "\n".join(compression.texts.get(hit.chunk.id, hit.chunk.text or "") for hit in budget.hits)
full_tokens.append(float((len(before_text) + 3) // 4))
compressed_tokens.append(float((len(after_text) + 3) // 4))
retention.append(float(_compression_signal(case, after_text)))
reductions.append(float(compression.reduction_ratio * 100.0))
return [
{
"configuration": "Adaptive context only",
"answer_signal_retention": 1.0,
"median_evidence_tokens_est": round(statistics.median(full_tokens), 1) if full_tokens else 0.0,
"median_additional_reduction_pct": 0.0,
"cases": len(qa_cases),
},
{
"configuration": "Adaptive + sentence compression",
"answer_signal_retention": round(mean(retention), 3),
"median_evidence_tokens_est": round(statistics.median(compressed_tokens), 1) if compressed_tokens else 0.0,
"median_additional_reduction_pct": round(statistics.median(reductions), 1) if reductions else 0.0,
"cases": len(qa_cases),
},
]
def _readiness_rows(summary: dict[str, Any], scale_rows: list[dict[str, Any]], compression_rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
level = str(summary.get("evaluation_level", "Standard"))
largest_scale = scale_rows[-1] if scale_rows else {}
compressed = compression_rows[-1] if compression_rows else {}
checks = [
("Answer accuracy", float(summary.get("answer_accuracy", 0.0)), 0.95, True),
("Source recall", float(summary.get("source_recall@5", 0.0)), 0.95, True),
("Citation validity", float(summary.get("citation_validity", 0.0)), 0.95, True),
("Citation coverage", float(summary.get("citation_coverage", 0.0)), 0.90, True),
("Planner route", float(summary.get("planner_route_accuracy", 0.0)), 0.90, True),
("Planner task", float(summary.get("planner_task_accuracy", 0.0)), 0.90, True),
("Planner strategy", float(summary.get("planner_strategy_accuracy", 0.0)), 0.90, True),
("Web precision", float(summary.get("web_use_precision", 0.0)), 0.90, True),
("Text2SQL", float(summary.get("text2sql_pass_rate", 0.0)), 0.90, True),
("Hard mode", float(summary.get("hard_mode_pass_rate", 0.0)), 0.85, True),
("Corpus overview", float(summary.get("overview_pass_rate", 0.0)), 0.90, True),
]
# Quick intentionally skips local ablations and scale stress. Do not award
# implicit PASS results for tests that were not executed. Standard/Deep
# require them and missing rows therefore fail the corresponding gate.
if level != "Quick":
checks.extend([
("Adaptive-budget recall", float(summary.get("context_pruning_recall@5", 0.0)), 0.95, True),
("Compression signal retention", float(compressed.get("answer_signal_retention", 0.0)), 0.95, True),
("20x stress recall", float(largest_scale.get("source_recall@5", 0.0)), 0.95, True),
("20x stress pruned recall", float(largest_scale.get("adaptive_pruned_recall@5", 0.0)), 0.95, True),
])
rows = []
for name, value, threshold, critical in checks:
rows.append({
"check": name,
"value": round(value, 3),
"threshold": threshold,
"status": "PASS" if value >= threshold else "FAIL",
"critical": critical,
})
return rows
def _readiness_summary(rows: list[dict[str, Any]]) -> tuple[str, float]:
critical = [row for row in rows if row.get("critical")]
passed = sum(1 for row in critical if row.get("status") == "PASS")
score = safe_div(passed, len(critical)) if critical else 0.0
if score >= 1.0:
return "READY", score
if score >= 0.85:
return "WATCH", score
return "NOT READY", score
def _planner_eval(
workspace: Workspace,
cases: list[dict[str, Any]],
gateway: GeminiGateway,
progress: Callable[[float, str], None],
start: float,
span: float,
) -> list[dict[str, Any]]:
manifest = workspace.manifest()
rows: list[dict[str, Any]] = []
total = max(1, len(cases))
for idx, case in enumerate(cases, start=1):
progress(start + span * (idx - 1) / total, f"Planner case {idx}/{len(cases)}")
wait_before = gateway.request_pacer.total_sleep_seconds() if gateway.request_pacer else 0.0
began = time.perf_counter()
plan = gateway.analyze_query(case["question"], manifest, history=None, profile="Balanced")
wall_latency = (time.perf_counter() - began) * 1000
wait_after = gateway.request_pacer.total_sleep_seconds() if gateway.request_pacer else wait_before
pacing_wait = max(0.0, wait_after - wait_before) * 1000
latency = max(0.0, wall_latency - pacing_wait)
planned_web = plan.web_relevance != "irrelevant" or plan.route in {"web", "hybrid"}
rows.append(
{
"id": case["id"],
"question": case["question"],
"expected_route": case["route"],
"route": plan.route,
"route_correct": plan.route == case["route"],
"expected_task": case["task"],
"task": plan.task_type,
"task_correct": plan.task_type == case["task"],
"expected_strategy": case["strategy"],
"strategy": plan.retrieval_strategy,
"strategy_correct": plan.retrieval_strategy in set(case.get("strategy_any", [case["strategy"]])),
"expected_web": bool(case["web_expected"]),
"planned_web": planned_web,
"latency_ms": round(latency, 1),
"wall_latency_ms": round(wall_latency, 1),
"pacing_wait_ms": round(pacing_wait, 1),
}
)
return rows
def _judge_row(
judge: GeminiGateway,
case: dict[str, Any],
answer: str,
sources: list[dict[str, Any]],
citations: dict[str, float | int],
) -> dict[str, Any]:
wait_before = judge.request_pacer.total_sleep_seconds() if judge.request_pacer else 0.0
began = time.perf_counter()
judgement = judge.evaluate_rag_answer(
case["question"],
answer,
_evidence_text(sources),
case.get("reference_answer", ""),
citation_validity=float(citations["citation_validity"]),
citation_coverage=float(citations["citation_coverage"]),
)
judge_wall_latency = (time.perf_counter() - began) * 1000
wait_after = judge.request_pacer.total_sleep_seconds() if judge.request_pacer else wait_before
judge_pacing_wait = max(0.0, wait_after - wait_before) * 1000
judge_latency = max(0.0, judge_wall_latency - judge_pacing_wait)
return {
"judge_faithfulness": round(judgement.faithfulness, 3),
"judge_answer_relevance": round(judgement.answer_relevance, 3),
"judge_completeness": round(judgement.completeness, 3),
"judge_citation_support": round(judgement.citation_support, 3),
"judge_overall": round(judgement.overall, 3),
"judge_pass": judgement.pass_,
"judge_reason": judgement.reason,
"judge_latency_ms": round(judge_latency, 1),
"judge_wall_latency_ms": round(judge_wall_latency, 1),
"judge_pacing_wait_ms": round(judge_pacing_wait, 1),
}
def _qa_eval(
workspace: Workspace,
cases: list[dict[str, Any]],
api_key: str | None,
model: str,
deep_judge: bool,
request_pacer: RequestPacer,
progress: Callable[[float, str], None],
start: float,
span: float,
) -> list[dict[str, Any]]:
engine = RAGEngine(workspace, request_pacer=request_pacer)
judge = GeminiGateway(api_key, model, request_pacer=request_pacer) if deep_judge else None
cfg = PipelineConfig(
mode="Documents",
profile="Fast",
model=model,
use_crag=False,
allow_web_fallback=False,
use_self_rag=False,
)
rows: list[dict[str, Any]] = []
total = max(1, len(cases))
for idx, case in enumerate(cases, start=1):
progress(start + span * (idx - 1) / total, f"Document QA case {idx}/{len(cases)}")
wait_before = request_pacer.total_sleep_seconds()
began = time.perf_counter()
result = engine.ask(case["question"], cfg, api_key, use_cache=False, record_history=False)
wall_latency = (time.perf_counter() - began) * 1000
pacing_wait = max(0.0, request_pacer.total_sleep_seconds() - wait_before) * 1000
latency = max(0.0, wall_latency - pacing_wait)
returned_sources = _document_sources(result.sources, 5)
retrieval = source_metrics(returned_sources, case.get("relevant_sources", []))
citations = citation_metrics(result.answer, result.sources)
efficiency = _trace_efficiency(result.trace)
row: dict[str, Any] = {
"id": case["id"],
"question": case["question"],
"answer_key_match": answer_key_match(result.answer, case),
**{key: round(float(value), 3) for key, value in retrieval.items()},
"citation_count": citations["citation_count"],
"citation_validity": round(float(citations["citation_validity"]), 3),
"citation_coverage": round(float(citations["citation_coverage"]), 3),
"confidence": round(result.confidence, 3),
"latency_ms": round(latency, 1),
"wall_latency_ms": round(wall_latency, 1),
"pacing_wait_ms": round(pacing_wait, 1),
**efficiency,
"_answer": result.answer,
"_sources": result.sources,
"_citations": citations,
"_node_times": _trace_node_times(result.trace, pacing_wait),
}
if judge and bool(case.get("deep_judge", False)):
row.update(_judge_row(judge, case, result.answer, result.sources, citations))
rows.append(row)
return rows
def _overview_eval(
workspace: Workspace,
cases: list[dict[str, Any]],
api_key: str | None,
model: str,
deep_judge: bool,
request_pacer: RequestPacer,
progress: Callable[[float, str], None],
start: float,
span: float,
) -> list[dict[str, Any]]:
engine = RAGEngine(workspace, request_pacer=request_pacer)
judge = GeminiGateway(api_key, model, request_pacer=request_pacer) if deep_judge else None
cfg = PipelineConfig(
mode="Auto",
profile="Balanced",
model=model,
allow_web_fallback=True,
use_crag=True,
use_self_rag=False,
)
rows: list[dict[str, Any]] = []
total = max(1, len(cases))
for idx, case in enumerate(cases, start=1):
progress(start + span * (idx - 1) / total, f"Corpus overview case {idx}/{len(cases)}")
wait_before = request_pacer.total_sleep_seconds()
began = time.perf_counter()
result = engine.ask(case["question"], cfg, api_key, use_cache=False, record_history=False)
wall_latency = (time.perf_counter() - began) * 1000
pacing_wait = max(0.0, request_pacer.total_sleep_seconds() - wait_before) * 1000
latency = max(0.0, wall_latency - pacing_wait)
plan = result.trace.get("query_plan", {})
evidence = result.trace.get("evidence", {})
efficiency = _trace_efficiency(result.trace)
coverage = float(evidence.get("source_coverage", 0.0) or 0.0)
citations = citation_metrics(result.answer, result.sources)
doc_sources = set(_document_sources(result.sources, 20))
expected_task = case["expected_task"]
expected_strategy = case["expected_strategy"]
actual_task = plan.get("task_type")
actual_strategy = plan.get("retrieval_strategy")
# A generic collection summary can legitimately be made richer as an
# insight synthesis. Treat overview/global and insight/analytical as the
# same broad-collection family for this suite, while still requiring
# local routing, breadth and no unnecessary web usage.
task_ok = actual_task == expected_task or (
expected_task == "overview" and actual_task == "insight_synthesis"
)
strategy_ok = actual_strategy == expected_strategy or (
expected_task == "overview"
and actual_task == "insight_synthesis"
and actual_strategy == "analytical"
)
passed = (
plan.get("route") == case["expected_route"]
and task_ok
and strategy_ok
and efficiency["web_used"] == bool(case["web_expected"])
and coverage >= float(case.get("min_source_coverage", 0.0))
and float(citations.get("citation_validity", 0.0)) >= 0.80
)
row: dict[str, Any] = {
"id": case["id"],
"question": case["question"],
"route": plan.get("route"),
"task": plan.get("task_type"),
"strategy": plan.get("retrieval_strategy"),
"task_semantic_match": task_ok,
"strategy_semantic_match": strategy_ok,
"web_used": efficiency["web_used"],
"source_coverage": round(coverage, 3),
"document_sources_returned": len(doc_sources),
"citation_validity": round(float(citations["citation_validity"]), 3),
"citation_coverage": round(float(citations["citation_coverage"]), 3),
"latency_ms": round(latency, 1),
"wall_latency_ms": round(wall_latency, 1),
"pacing_wait_ms": round(pacing_wait, 1),
"pass": passed,
**efficiency,
"_answer": result.answer,
"_sources": result.sources,
"_citations": citations,
"_node_times": _trace_node_times(result.trace, pacing_wait),
}
if judge and bool(case.get("deep_judge", False)):
row.update(_judge_row(judge, case, result.answer, result.sources, citations))
rows.append(row)
return rows
def _sql_eval(
workspace: Workspace,
cases: list[dict[str, Any]],
api_key: str | None,
model: str,
request_pacer: RequestPacer,
progress: Callable[[float, str], None],
start: float,
span: float,
) -> list[dict[str, Any]]:
"""Evaluate Text2SQL generation/execution with one model call per case.
SQL routing is already measured in the semantic-planner benchmark. Keeping
this component test route-independent avoids spending two extra Gemini
calls per case just to duplicate planner and answer-generation coverage.
"""
gateway = GeminiGateway(api_key, model, request_pacer=request_pacer)
rows: list[dict[str, Any]] = []
total = max(1, len(cases))
for idx, case in enumerate(cases, start=1):
progress(start + span * (idx - 1) / total, f"Text2SQL case {idx}/{len(cases)}")
wait_before = request_pacer.total_sleep_seconds()
began = time.perf_counter()
try:
sql, result = workspace.sql.benchmark_query(case["question"], gateway)
preview = result.head(200)
result_text = preview.to_markdown(index=False) if len(preview) else "(no rows)"
observed_scalar = preview.iloc[0, 0] if len(preview) and len(preview.columns) else None
if "expected_scalar" in case:
matched = scalar_value_match(observed_scalar, case.get("expected_scalar"))
match_method = "typed_scalar"
else:
matched = answer_key_match(result_text, case)
match_method = "rendered_answer_key"
error = ""
readonly_validated = True
except Exception as exc:
sql = ""
result = None
result_text = ""
matched = False
match_method = "error"
observed_scalar = None
error = f"{type(exc).__name__}: {exc}"
readonly_validated = False
wall_latency = (time.perf_counter() - began) * 1000
pacing_wait = max(0.0, request_pacer.total_sleep_seconds() - wait_before) * 1000
latency = max(0.0, wall_latency - pacing_wait)
rows.append(
{
"id": case["id"],
"question": case["question"],
"component": "Text2SQL",
"answer_key_match": matched,
"match_method": match_method,
"observed_value": (
observed_scalar.item() if hasattr(observed_scalar, "item") else observed_scalar
),
"expected_value": case.get("expected_scalar", ""),
"readonly_validated": readonly_validated,
"sql": sql,
"rows": int(len(result)) if result is not None else 0,
"latency_ms": round(latency, 1),
"wall_latency_ms": round(wall_latency, 1),
"pacing_wait_ms": round(pacing_wait, 1),
"llm_calls_estimate": 1,
"error": error,
}
)
return rows
def _abstention_eval() -> list[dict[str, Any]]:
empty = Workspace(f"eval-empty-{uuid.uuid4().hex[:10]}")
engine = RAGEngine(empty)
cases = [
("Documents", "What are the documents about?", "workspace_empty_documents"),
("Data (SQL)", "Which row has the highest value?", "workspace_empty_tables"),
]
rows = []
for mode, question, expected_reason in cases:
result = engine.ask(
question,
PipelineConfig(mode=mode, profile="Fast"),
api_key=None,
use_cache=False,
record_history=False,
)
nodes = result.trace.get("nodes", [])
abstain_node = next((node for node in nodes if node.get("node") == "abstain"), {})
rows.append(
{
"mode": mode,
"question": question,
"abstained": bool(abstain_node),
"reason": abstain_node.get("reason"),
"expected_reason": expected_reason,
"pass": abstain_node.get("reason") == expected_reason,
"llm_calls_estimate": result.trace.get("metrics", {}).get("llm_calls_estimate", 0),
}
)
return rows
def _planner_summary(rows: list[dict[str, Any]]) -> dict[str, float]:
tp = fp = fn = 0
for row in rows:
expected = bool(row["expected_web"])
planned = bool(row["planned_web"])
if expected and planned:
tp += 1
elif not expected and planned:
fp += 1
elif expected and not planned:
fn += 1
return {
"planner_route_accuracy": mean([float(row["route_correct"]) for row in rows]),
"planner_task_accuracy": mean([float(row["task_correct"]) for row in rows]),
"planner_strategy_accuracy": mean([float(row["strategy_correct"]) for row in rows]),
"web_use_precision": safe_div(tp, tp + fp),
"web_use_recall": safe_div(tp, tp + fn),
"unnecessary_web_rate": safe_div(fp, sum(1 for row in rows if not row["expected_web"])),
}
def _base_grade(score: float) -> str:
if score >= 0.90:
return "A"
if score >= 0.80:
return "B"
if score >= 0.70:
return "C"
if score >= 0.60:
return "D"
return "Needs work"
def _grade_with_gates(score: float, metrics: dict[str, float]) -> tuple[str, list[str]]:
"""Prevent a weighted average from hiding a badly failing subsystem."""
order = ["Needs work", "D", "C", "B", "A"]
grade = _base_grade(score)
gates: list[str] = []
def cap(max_grade: str, reason: str) -> None:
nonlocal grade
if order.index(grade) > order.index(max_grade):
grade = max_grade
gates.append(reason)
if metrics["planner_route_accuracy"] < 0.90 or metrics["web_use_precision"] < 0.90:
cap("C", "Critical routing/web-policy accuracy is below 90%.")
if metrics["text2sql_pass_rate"] < 0.75:
cap("B", "Text2SQL pass rate is below 75%.")
if metrics["citation_validity"] < 0.90:
cap("B", "Citation validity is below 90%.")
if metrics["citation_coverage"] < 0.80:
cap("B", "Citation coverage is below 80%.")
if metrics["planner_task_accuracy"] < 0.75:
cap("B", "Planner task taxonomy accuracy is below 75%.")
if metrics.get("hard_mode_pass_rate", 1.0) < 0.50:
cap("C", "Hard-mode robustness pass rate is below 50%.")
elif metrics.get("hard_mode_pass_rate", 1.0) < 0.75:
cap("B", "Hard-mode robustness pass rate is below 75%.")
return grade, gates
def _hard_mode_eval(
workspace: Workspace,
cases: list[dict[str, Any]],
api_key: str | None,
model: str,
request_pacer: RequestPacer,
progress: Callable[[float, str], None],
start: float,
span: float,
) -> list[dict[str, Any]]:
engine = RAGEngine(workspace, request_pacer=request_pacer)
gateway = GeminiGateway(api_key, model, request_pacer=request_pacer)
rows: list[dict[str, Any]] = []
total = max(1, len(cases))
for idx, case in enumerate(cases, start=1):
progress(start + span * (idx - 1) / total, f"Hard-mode case {idx}/{len(cases)}")
kind = case.get("kind", "qa")
row: dict[str, Any] = {"id": case["id"], "kind": kind, "question": case.get("question", "")}
if kind == "security":
score = prompt_injection_score(case.get("text", ""))
row.update({"injection_score": round(score, 3), "pass": score >= float(case.get("min_injection_score", 0.5)), "gemini_calls": 0})
rows.append(row)
continue
if kind == "sql":
wait_before = request_pacer.total_sleep_seconds()
began = time.perf_counter()
try:
sql, result = workspace.sql.benchmark_query(case["question"], gateway)
observed = result.iloc[0, 0] if len(result) and len(result.columns) else None
passed = scalar_value_match(observed, case.get("expected_scalar"))
error = ""
except Exception as exc:
sql, observed, passed = "", None, False
error = f"{type(exc).__name__}: {exc}"
wall = (time.perf_counter() - began) * 1000
pace = max(0.0, request_pacer.total_sleep_seconds() - wait_before) * 1000
row.update({"observed_value": observed.item() if hasattr(observed, "item") else observed, "expected_value": case.get("expected_scalar"), "sql": sql, "pass": passed, "latency_ms": round(max(0.0, wall-pace),1), "error": error, "gemini_calls": 1})
rows.append(row)
continue
if kind == "planner":
wait_before = request_pacer.total_sleep_seconds()
began = time.perf_counter()
plan = gateway.analyze_query(case["question"], workspace.manifest(), history=None, profile="Balanced")
wall = (time.perf_counter() - began) * 1000
pace = max(0.0, request_pacer.total_sleep_seconds() - wait_before) * 1000
passed = plan.route == case["route"] and plan.task_type == case["task"] and plan.retrieval_strategy == case["strategy"] and ((plan.web_relevance != "irrelevant") == bool(case["web_expected"]))
row.update({"route":plan.route,"task":plan.task_type,"strategy":plan.retrieval_strategy,"web_relevance":plan.web_relevance,"pass":passed,"latency_ms":round(max(0.0,wall-pace),1),"gemini_calls":1})
rows.append(row)
continue
case_mode = str(case.get("mode") or "Documents")
case_profile = str(case.get("profile") or "Fast")
cfg = PipelineConfig(
mode=case_mode,
profile=case_profile,
model=model,
use_crag=bool(case.get("use_crag", False)),
allow_web_fallback=False,
use_self_rag=False,
)
if kind == "insight":
cfg = PipelineConfig(mode="Auto", profile="Balanced", model=model, use_crag=True, allow_web_fallback=False, use_self_rag=False)
wait_before = request_pacer.total_sleep_seconds()
began = time.perf_counter()
result = engine.ask(case["question"], cfg, api_key, use_cache=False, record_history=False)
wall = (time.perf_counter() - began) * 1000
pace = max(0.0, request_pacer.total_sleep_seconds() - wait_before) * 1000
plan = result.trace.get("query_plan", {})
citations = citation_metrics(result.answer, result.sources)
returned = _document_sources(result.sources, 5)
retrieval = source_metrics(returned, case.get("relevant_sources", [])) if case.get("relevant_sources") else {}
if kind == "missing":
grounded_absence = bool(result.trace.get("metrics", {}).get("grounded_absence", False))
passed = missing_answer_match(result.answer, case) or grounded_absence
elif kind == "insight":
table_cited = any(str(src.get("id", "")).startswith("T") for src in result.sources) and "[T" in result.answer
evidence = result.trace.get("evidence", {})
passed = plan.get("task_type") == case.get("expected_task") and plan.get("retrieval_strategy") == case.get("expected_strategy") and float(evidence.get("source_coverage", 0.0) or 0.0) >= float(case.get("min_source_coverage", 0.0)) and (table_cited or not case.get("requires_table_citation"))
else:
passed = answer_key_match(result.answer, case) and float(retrieval.get("source_recall@5", 1.0)) >= 1.0
expected_route = case.get("expected_route")
expected_task = case.get("expected_task")
expected_strategy = case.get("expected_strategy")
if expected_route:
passed = passed and plan.get("route") == expected_route
if expected_task:
passed = passed and plan.get("task_type") == expected_task
if expected_strategy:
passed = passed and plan.get("retrieval_strategy") == expected_strategy
row.update({
"evaluation_mode": cfg.mode if kind not in {"sql", "planner", "security"} else None,
"evaluation_profile": cfg.profile if kind not in {"sql", "planner", "security"} else None,
"route": plan.get("route"),
"task": plan.get("task_type"),
"strategy": plan.get("retrieval_strategy"),
"answer_key_match": answer_key_match(result.answer, case) if kind == "qa" else None,
"missing_answer_match": missing_answer_match(result.answer, case) if kind == "missing" else None,
"grounded_absence": bool(result.trace.get("metrics", {}).get("grounded_absence", False)) if kind == "missing" else None,
"citation_validity": round(float(citations["citation_validity"]), 3),
"citation_coverage": round(float(citations["citation_coverage"]), 3),
"source_recall@5": round(float(retrieval.get("source_recall@5", 1.0)), 3),
"latency_ms": round(max(0.0, wall - pace), 1),
"pass": passed,
"gemini_calls": int(result.trace.get("metrics", {}).get("llm_calls_estimate", 0) or 0),
"_answer": result.answer,
"_sources": result.sources,
"_node_times": _trace_node_times(result.trace, pace),
})
rows.append(row)
return rows
def _profile_benchmark(
workspace: Workspace,
cases: list[dict[str, Any]],
api_key: str | None,
model: str,
request_pacer: RequestPacer,
progress: Callable[[float, str], None],
start: float,
span: float,
) -> list[dict[str, Any]]:
selected = [cases[0]] if cases else []
cross = next((case for case in cases if len(case.get("relevant_sources", [])) > 1), None)
if cross and cross not in selected:
selected.append(cross)
elif len(cases) > 1:
selected.append(cases[1])
rows: list[dict[str, Any]] = []
combinations = [(profile, case) for profile in ("Fast", "Balanced", "Agentic") for case in selected]
total = max(1, len(combinations))
for idx, (profile, case) in enumerate(combinations, start=1):
progress(start + span * (idx-1)/total, f"Profile benchmark {idx}/{len(combinations)}")
cfg = PipelineConfig(mode="Documents", profile=profile, model=model, allow_web_fallback=False, use_crag=True, use_self_rag=True)
engine = RAGEngine(workspace, request_pacer=request_pacer)
wait_before=request_pacer.total_sleep_seconds(); began=time.perf_counter()
result=engine.ask(case["question"],cfg,api_key,use_cache=False,record_history=False)
wall=(time.perf_counter()-began)*1000; pace=max(0.0,request_pacer.total_sleep_seconds()-wait_before)*1000
citations=citation_metrics(result.answer,result.sources); metrics=result.trace.get("metrics",{})
rows.append({"profile":profile,"case":case["id"],"answer_key_match":answer_key_match(result.answer,case),"citation_validity":round(float(citations["citation_validity"]),3),"citation_coverage":round(float(citations["citation_coverage"]),3),"latency_ms":round(max(0.0,wall-pace),1),"llm_calls_estimate":int(metrics.get("llm_calls_estimate",0) or 0),"reranker_used":bool(metrics.get("reranker_used",False)),"correction_used":bool(metrics.get("correction_used",False))})
return rows
def _profile_summary(rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
summaries: list[dict[str, Any]] = []
for profile in ("Fast", "Balanced", "Agentic"):
group = [row for row in rows if row.get("profile") == profile]
if not group:
continue
summaries.append({
"profile": profile,
"answer_accuracy": round(mean([float(bool(row.get("answer_key_match"))) for row in group]), 3),
"citation_validity": round(mean([float(row.get("citation_validity", 0.0) or 0.0) for row in group]), 3),
"citation_coverage": round(mean([float(row.get("citation_coverage", 0.0) or 0.0) for row in group]), 3),
"median_latency_ms": round(percentile([float(row.get("latency_ms", 0.0) or 0.0) for row in group], 0.5), 1),
"mean_llm_calls": round(mean([float(row.get("llm_calls_estimate", 0.0) or 0.0) for row in group]), 2),
"reranker_rate": round(mean([float(bool(row.get("reranker_used"))) for row in group]), 3),
"cases": len(group),
})
return summaries
def _profile_recommendation(profile_summary: list[dict[str, Any]]) -> str:
if not profile_summary:
return ""
by_name = {row["profile"]: row for row in profile_summary}
fast = by_name.get("Fast")
balanced = by_name.get("Balanced")
agentic = by_name.get("Agentic")
if fast and balanced:
quality_close = (
float(fast.get("answer_accuracy", 0.0)) >= float(balanced.get("answer_accuracy", 0.0)) - 0.01
and float(fast.get("citation_coverage", 0.0)) >= float(balanced.get("citation_coverage", 0.0)) - 0.05
)
faster = float(fast.get("median_latency_ms", 0.0) or 0.0) < float(balanced.get("median_latency_ms", 0.0) or 0.0)
if quality_close and faster:
return "Fast matched Balanced quality on the sampled explicit-Documents cases with lower median latency. Keep Balanced as the general Auto default, but prefer Fast for simple local lookups."
if agentic and balanced and float(agentic.get("median_latency_ms", 0.0) or 0.0) > 2 * max(1.0, float(balanced.get("median_latency_ms", 0.0) or 0.0)):
return "Agentic was materially slower than Balanced on the sampled cases. Reserve Agentic for difficult or low-confidence work rather than routine lookups."
return "Profile differences were not large enough on this sample to justify changing the default execution policy."
def _diagnostics(
summary: dict[str, Any],
ablation_rows: list[dict[str, Any]],
planner_rows: list[dict[str, Any]],
sql_rows: list[dict[str, Any]],
hard_rows: list[dict[str, Any]] | None = None,
profile_summary: list[dict[str, Any]] | None = None,
node_latency_rows: list[dict[str, Any]] | None = None,
context_budget_rows: list[dict[str, Any]] | None = None,
compression_rows: list[dict[str, Any]] | None = None,
scale_stress_rows: list[dict[str, Any]] | None = None,
) -> list[dict[str, str]]:
findings: list[dict[str, str]] = []
if int(summary.get("rate_limit_retries", 0) or 0) > 0:
findings.append(
{
"severity": "warning",
"area": "gemini quota",
"finding": (
f"Gemini surfaced {int(summary.get('rate_limit_retries', 0))} rate-limit retry event(s); "
f"provider-directed retry wait was {float(summary.get('rate_limit_sleep_ms', 0.0)) / 1000:.1f}s."
),
"recommendation": "Keep quota-safe pacing enabled or lower the target RPM below the active project limit.",
}
)
context_budget_rows = context_budget_rows or []
adaptive_row = next(
(row for row in context_budget_rows if row.get("configuration") == "Adaptive budget"), {}
)
full_row = next((row for row in context_budget_rows if row.get("configuration") == "Full top-k"), {})
if summary.get("source_recall@5", 0.0) >= 0.95 and summary.get("source_precision@5", 1.0) < 0.60:
if adaptive_row:
findings.append({
"severity": "info",
"area": "context efficiency",
"finding": (
f"Source Recall@5 is {float(summary.get('source_recall@5', 0.0)):.0%}; adaptive focused context now "
f"uses a median target of {float(adaptive_row.get('median_target_chunks', 0.0)):.1f} chunks and reduces "
f"context by {float(adaptive_row.get('median_context_reduction_pct', 0.0)):.0f}%."
),
"recommendation": "The runtime already applies adaptive budgeting. Use the compression and scale-stress ablations to decide whether further tightening is safe rather than lowering top-k globally.",
})
else:
findings.append({
"severity": "info",
"area": "context efficiency",
"finding": (
f"Source Recall@5 is {float(summary.get('source_recall@5', 0.0)):.0%} while source Precision@5 is "
f"{float(summary.get('source_precision@5', 0.0)):.0%}."
),
"recommendation": "Use a focused context budget before reducing global retrieval breadth.",
})
if adaptive_row and full_row:
full_precision = float(full_row.get("source_precision@5", 0.0) or 0.0)
adaptive_precision = float(adaptive_row.get("source_precision@5", 0.0) or 0.0)
full_recall = float(full_row.get("source_recall@5", 0.0) or 0.0)
adaptive_recall = float(adaptive_row.get("source_recall@5", 0.0) or 0.0)
reduction = float(adaptive_row.get("median_context_reduction_pct", 0.0) or 0.0)
if adaptive_recall >= full_recall - 1e-9 and reduction >= 25.0:
findings.append({
"severity": "ok",
"area": "adaptive context budget",
"finding": (
f"Adaptive budgeting preserved source Recall@5 at {adaptive_recall:.0%}, changed Precision@5 from "
f"{full_precision:.0%} to {adaptive_precision:.0%}, and cut median context by {reduction:.0f}%."
),
"recommendation": "Keep adaptive budgeting enabled; corpus-scale stress now provides the guardrail for future budget changes.",
})
elif adaptive_recall < full_recall - 1e-9:
findings.append({
"severity": "warning",
"area": "adaptive context budget",
"finding": f"Adaptive budgeting reduced source Recall@5 from {full_recall:.0%} to {adaptive_recall:.0%}.",
"recommendation": "Loosen the adaptive budget before shipping this policy broadly.",
})
compression_rows = compression_rows or []
compressed = next(
(row for row in compression_rows if row.get("configuration") == "Adaptive + sentence compression"), {}
)
if compressed:
retention = float(compressed.get("answer_signal_retention", 0.0) or 0.0)
reduction = float(compressed.get("median_additional_reduction_pct", 0.0) or 0.0)
findings.append({
"severity": "ok" if retention >= 0.95 else "warning",
"area": "evidence compression",
"finding": f"Focused sentence compression retained labeled answer signals in {retention:.0%} of cases while cutting selected-evidence tokens by a median {reduction:.0f}% beyond context budgeting.",
"recommendation": "Keep compression enabled for focused lookups only." if retention >= 0.95 else "Disable or loosen sentence compression until labeled signal retention returns above 95%.",
})
scale_stress_rows = scale_stress_rows or []
if summary.get("scale_stress_error"):
findings.append({
"severity": "warning",
"area": "scale stress",
"finding": f"The zero-Gemini scale-stress harness did not complete: {summary.get('scale_stress_error')}",
"recommendation": "Treat release readiness as incomplete until the local scale-stress harness runs successfully; the main RAG benchmark can still be inspected independently.",
})
if scale_stress_rows:
largest = scale_stress_rows[-1]
recall = float(largest.get("source_recall@5", 0.0) or 0.0)
pruned_recall = float(largest.get("adaptive_pruned_recall@5", 0.0) or 0.0)
findings.append({
"severity": "ok" if min(recall, pruned_recall) >= 0.95 else "warning",
"area": "scale stress",
"finding": (
f"At {int(largest.get('chunks', 0) or 0)} chunks / {int(largest.get('sources', 0) or 0)} sources, "
f"retrieval Recall@5 was {recall:.0%} and adaptive-pruned recall was {pruned_recall:.0%}."
),
"recommendation": "Treat this as synthetic distractor evidence, then repeat with a real larger upload before changing the reranker policy." if min(recall, pruned_recall) >= 0.95 else "Increase retrieval depth or budget targets for large corpora before relying on the adaptive policy.",
})
if summary.get("citation_coverage", 1.0) < 0.90:
findings.append(
{
"severity": "warning",
"area": "citations",
"finding": f"Citation coverage is {float(summary['citation_coverage']):.0%}; some factual statements are uncited.",
"recommendation": "Keep the generation prompt citation requirement and inspect low-coverage cases individually.",
}
)
if summary.get("planner_task_accuracy", 1.0) < 0.90:
failed = [row["id"] for row in planner_rows if not row.get("task_correct")]
findings.append(
{
"severity": "warning",
"area": "planner",
"finding": f"Task classification misses: {', '.join(failed) or 'none'}.",
"recommendation": "Review task taxonomy labels separately from route/strategy correctness; do not over-penalize equivalent plans.",
}
)
if summary.get("text2sql_pass_rate", 1.0) < 0.90:
failed = [row["id"] for row in sql_rows if not row.get("answer_key_match")]
findings.append(
{
"severity": "warning",
"area": "text2sql",
"finding": f"Text2SQL failed cases: {', '.join(failed) or 'none'}.",
"recommendation": (
"Inspect generated SQL, typed observed values and benchmark expectations. SQL routing is evaluated "
"separately in the planner suite, so a component failure should not automatically be blamed on routing."
),
}
)
hard_rows = hard_rows or []
if hard_rows and summary.get("hard_mode_pass_rate", 1.0) < 0.90:
failed = [row.get("id", "?") for row in hard_rows if not row.get("pass")]
findings.append({
"severity": "warning",
"area": "hard_mode",
"finding": f"Hard-mode robustness failures: {', '.join(failed) or 'none'}.",
"recommendation": "Inspect missing-answer, distractor, analytical and adversarial cases before expanding the feature set.",
})
if len(ablation_rows) == 2:
base, rerank = ablation_rows
base_ms = float(base.get("median_retrieval_ms", 0.0) or 0.0)
rerank_ms = float(rerank.get("median_retrieval_ms", 0.0) or 0.0)
base_mrr = float(base.get("source_mrr", 0.0) or 0.0)
rerank_mrr = float(rerank.get("source_mrr", 0.0) or 0.0)
multiplier = safe_div(rerank_ms, base_ms) if base_ms else 0.0
if multiplier >= 3.0 and rerank_mrr <= base_mrr + 0.01:
findings.append(
{
"severity": "info",
"area": "reranker",
"finding": f"The reranker ablation is {multiplier:.1f}x slower on the demo benchmark with no material source-MRR gain.",
"recommendation": "The runtime already skips reranking on small corpora. Keep this ablation as evidence and re-enable the cross-encoder only when a larger-corpus benchmark shows source- or chunk-level gain.",
}
)
profile_summary = profile_summary or []
recommendation = _profile_recommendation(profile_summary)
if recommendation:
findings.append({
"severity": "info",
"area": "profile policy",
"finding": recommendation,
"recommendation": "Use the profile benchmark as local evidence only; repeat it on larger user corpora before making a global policy claim.",
})
node_latency_rows = node_latency_rows or []
if node_latency_rows:
dominant = max(node_latency_rows, key=lambda row: float(row.get("mean_ms", 0.0) or 0.0))
total_mean = sum(float(row.get("mean_ms", 0.0) or 0.0) for row in node_latency_rows)
share = safe_div(float(dominant.get("mean_ms", 0.0) or 0.0), total_mean)
if share >= 0.60:
findings.append({
"severity": "info",
"area": "latency",
"finding": f"{dominant.get('node', 'generation')} dominates mean node time at approximately {share:.0%} of measured pipeline-node latency.",
"recommendation": "Prioritize model/generation efficiency before micro-optimizing millisecond-scale retrieval stages.",
})
if not findings:
findings.append(
{
"severity": "ok",
"area": "benchmark",
"finding": "No configured quality gate produced a diagnostic warning.",
"recommendation": "Expand the benchmark before treating this as general performance evidence.",
}
)
return findings
def _deep_from_standard_cache(
base_report: dict[str, Any],
api_key: str | None,
model: str,
*,
target_rpm: int,
progress: Callable[[float, str], None],
) -> dict[str, Any]:
"""Upgrade a cached Standard run to Deep with judge calls only.
The deterministic suites are identical between Standard and Deep. Reusing a
current Standard baseline avoids spending ~25 repeated Gemini calls merely
to regenerate metrics the user already computed. Deep then adds the sampled
calibrated judge layer on top of those exact answers/evidence artifacts.
"""
started = time.perf_counter()
report = json.loads(json.dumps(base_report))
benchmark = _load_benchmark()
cases = {
case["id"]: case
for case in benchmark.get("qa_cases", []) + benchmark.get("overview_cases", [])
if case.get("deep_judge")
}
request_pacer = RequestPacer(target_rpm=max(0, int(target_rpm)))
judge = GeminiGateway(api_key, model, request_pacer=request_pacer)
judge_rows: list[dict[str, Any]] = []
candidate_rows: list[dict[str, Any]] = []
for section in ("focused_qa", "corpus_overviews"):
for row in report.get(section, []):
if row.get("id") in cases and row.get("_answer") is not None and row.get("_sources") is not None:
candidate_rows.append(row)
total = max(1, len(candidate_rows))
for idx, row in enumerate(candidate_rows, start=1):
progress(0.08 + 0.84 * (idx - 1) / total, f"Deep judge case {idx}/{len(candidate_rows)}")
case = cases[row["id"]]
citations = row.get("_citations") or citation_metrics(row.get("_answer", ""), row.get("_sources", []))
row.update(_judge_row(judge, case, row.get("_answer", ""), row.get("_sources", []), citations))
judge_rows.append(row)
summary = report.setdefault("summary", {})
baseline_wall_ms = float(summary.get("evaluation_wall_ms", 0.0) or 0.0)
if judge_rows:
summary.update(
{
"judge_faithfulness": round(mean([float(row["judge_faithfulness"]) for row in judge_rows]), 3),
"judge_answer_relevance": round(mean([float(row["judge_answer_relevance"]) for row in judge_rows]), 3),
"judge_completeness": round(mean([float(row["judge_completeness"]) for row in judge_rows]), 3),
"judge_citation_support": round(mean([float(row["judge_citation_support"]) for row in judge_rows]), 3),
"judge_overall": round(mean([float(row["judge_overall"]) for row in judge_rows]), 3),
"judge_pass_rate": round(mean([float(row["judge_pass"]) for row in judge_rows]), 3),
"judge_latency_mean_ms": round(mean([float(row.get("judge_latency_ms", 0.0)) for row in judge_rows]), 3),
}
)
pacing_stats = request_pacer.stats()
summary.update(
{
"evaluation_level": "Deep",
"evaluation_wall_ms": round((time.perf_counter() - started) * 1000, 1),
"evaluation_target_rpm": int(pacing_stats["target_rpm"]),
"gemini_requests": int(pacing_stats["gemini_requests"]),
"pacing_sleep_ms": float(pacing_stats["pacing_sleep_ms"]),
"rate_limit_retries": int(pacing_stats["rate_limit_retries"]),
"rate_limit_sleep_ms": float(pacing_stats["rate_limit_sleep_ms"]),
"deep_judge_cases": len(judge_rows),
"reused_standard_baseline": True,
"deep_incremental": True,
"deterministic_baseline_wall_ms": round(baseline_wall_ms, 1),
}
)
report["diagnostics"] = _diagnostics(
summary,
report.get("retrieval_ablation", []),
report.get("semantic_planner", []),
report.get("text2sql", []),
report.get("hard_mode", []),
report.get("profile_summary", []),
report.get("node_latency", []),
report.get("context_budget_ablation", []),
report.get("evidence_compression_ablation", []),
report.get("scale_stress", []),
)
report.setdefault("methodology", {})["evaluation_cache"] = (
"Deep reused the current cached Standard deterministic baseline and issued only sampled judge calls."
)
progress(1.0, "Deep evaluation complete")
return report
def run_demo_eval(
workspace: Workspace,
api_key: str | None,
model: str,
level: str = "Standard",
progress_callback: Callable[[float, str], None] | None = None,
target_rpm: int = 12,
base_standard_report: dict[str, Any] | None = None,
include_profile_benchmark: bool = False,
) -> dict[str, Any]:
"""Run the bundled benchmark with response caching disabled.
Quick: smaller deterministic regression set.
Standard: full deterministic set + retrieval and context-budget ablations.
Deep: Standard plus calibrated Gemini LLM-as-judge scores.
"""
wall_started = time.perf_counter()
progress = progress_callback or (lambda _value, _message: None)
benchmark = _load_benchmark()
level = level if level in {"Quick", "Standard", "Deep"} else "Standard"
if level == "Deep" and base_standard_report and not include_profile_benchmark:
artifact_rows = base_standard_report.get("focused_qa", []) + base_standard_report.get("corpus_overviews", [])
if any(row.get("_answer") is not None and row.get("_sources") is not None for row in artifact_rows):
return _deep_from_standard_cache(
base_standard_report,
api_key,
model,
target_rpm=target_rpm,
progress=progress,
)
deep_judge = level == "Deep"
request_pacer = RequestPacer(target_rpm=max(0, int(target_rpm)))
qa_cases = benchmark["qa_cases"] if level != "Quick" else benchmark["qa_cases"][:3]
planner_cases = benchmark["planner_cases"] if level != "Quick" else benchmark["planner_cases"][:5]
overview_cases = benchmark["overview_cases"] if level != "Quick" else benchmark["overview_cases"][:1]
sql_cases = benchmark.get("sql_cases", []) if level != "Quick" else benchmark.get("sql_cases", [])[:1]
hard_cases = benchmark.get("hard_mode_cases", []) if level != "Quick" else benchmark.get("hard_mode_cases", [])[:2]
progress(0.01, "Preparing evaluation")
qa_rows = _qa_eval(
workspace, qa_cases, api_key, model, deep_judge, request_pacer, progress, 0.03, 0.29
)
gateway = GeminiGateway(api_key, model, request_pacer=request_pacer)
planner_rows = _planner_eval(workspace, planner_cases, gateway, progress, 0.34, 0.22)
overview_rows = _overview_eval(
workspace, overview_cases, api_key, model, deep_judge, request_pacer, progress, 0.58, 0.16
)
sql_rows = (
_sql_eval(workspace, sql_cases, api_key, model, request_pacer, progress, 0.75, 0.08)
if sql_cases
else []
)
hard_rows = _hard_mode_eval(
workspace, hard_cases, api_key, model, request_pacer, progress, 0.83, 0.10
) if hard_cases else []
profile_rows = _profile_benchmark(
workspace, qa_cases, api_key, model, request_pacer, progress, 0.93, 0.05
) if include_profile_benchmark and level != "Quick" else []
progress(0.98, "Checking abstention, retrieval, context and scale ablations")
abstention_rows = _abstention_eval()
ablation_rows = _retrieval_ablation(workspace, qa_cases) if level != "Quick" else []
context_budget_rows = _context_budget_ablation(workspace, qa_cases) if level != "Quick" else []
compression_rows = _evidence_compression_ablation(workspace, qa_cases) if level != "Quick" else []
scale_stress_error = ""
if level != "Quick":
try:
scale_stress_rows = scale_stress_retrieval_eval(workspace, qa_cases)
except Exception as exc: # keep the primary benchmark available even if the local stress harness fails
scale_stress_rows = []
scale_stress_error = f"{type(exc).__name__}: {exc}"
else:
scale_stress_rows = []
planner_metrics = _planner_summary(planner_rows)
qa_latencies = [float(row["latency_ms"]) for row in qa_rows]
overview_latencies = [float(row["latency_ms"]) for row in overview_rows]
sql_latencies = [float(row["latency_ms"]) for row in sql_rows]
planner_latencies = [float(row["latency_ms"]) for row in planner_rows]
all_latencies = qa_latencies + overview_latencies + sql_latencies
all_runtime_rows = qa_rows + overview_rows + sql_rows + [row for row in hard_rows if row.get("latency_ms") is not None]
answer_accuracy = mean([float(row["answer_key_match"]) for row in qa_rows])
source_recall = mean([float(row["source_recall@5"]) for row in qa_rows])
source_mrr = mean([float(row["source_mrr"]) for row in qa_rows])
citation_validity = mean([float(row["citation_validity"]) for row in qa_rows + overview_rows])
citation_coverage = mean([float(row["citation_coverage"]) for row in qa_rows + overview_rows])
overview_pass = mean([float(row["pass"]) for row in overview_rows])
abstention_accuracy = mean([float(row["pass"]) for row in abstention_rows])
sql_accuracy = mean([float(row["answer_key_match"]) for row in sql_rows]) if sql_rows else 1.0
hard_accuracy = mean([float(row.get("pass", False)) for row in hard_rows]) if hard_rows else 1.0
deterministic_score = (
0.19 * answer_accuracy
+ 0.12 * source_recall
+ 0.06 * source_mrr
+ 0.08 * citation_validity
+ 0.07 * citation_coverage
+ 0.11 * planner_metrics["planner_route_accuracy"]
+ 0.07 * planner_metrics["planner_task_accuracy"]
+ 0.07 * planner_metrics["planner_strategy_accuracy"]
+ 0.07 * planner_metrics["web_use_precision"]
+ 0.04 * overview_pass
+ 0.04 * abstention_accuracy
+ 0.03 * sql_accuracy
+ 0.05 * hard_accuracy
)
judge_rows = [row for row in qa_rows + overview_rows if "judge_overall" in row]
judge_summary: dict[str, float] = {}
if judge_rows:
judge_summary = {
"judge_faithfulness": mean([float(row["judge_faithfulness"]) for row in judge_rows]),
"judge_answer_relevance": mean([float(row["judge_answer_relevance"]) for row in judge_rows]),
"judge_completeness": mean([float(row["judge_completeness"]) for row in judge_rows]),
"judge_citation_support": mean([float(row["judge_citation_support"]) for row in judge_rows]),
"judge_overall": mean([float(row["judge_overall"]) for row in judge_rows]),
"judge_pass_rate": mean([float(row["judge_pass"]) for row in judge_rows]),
"judge_latency_mean_ms": mean([float(row.get("judge_latency_ms", 0.0)) for row in judge_rows]),
}
pacing_stats = request_pacer.stats()
metrics_for_gate = {
"planner_route_accuracy": planner_metrics["planner_route_accuracy"],
"web_use_precision": planner_metrics["web_use_precision"],
"planner_task_accuracy": planner_metrics["planner_task_accuracy"],
"citation_validity": citation_validity,
"citation_coverage": citation_coverage,
"text2sql_pass_rate": sql_accuracy,
"hard_mode_pass_rate": hard_accuracy,
}
grade, quality_gates = _grade_with_gates(deterministic_score, metrics_for_gate)
adaptive_budget_row = next(
(row for row in context_budget_rows if row.get("configuration") == "Adaptive budget"), {}
)
compression_row = next(
(row for row in compression_rows if row.get("configuration") == "Adaptive + sentence compression"), {}
)
largest_scale_row = scale_stress_rows[-1] if scale_stress_rows else {}
summary: dict[str, Any] = {
"benchmark_version": benchmark.get("version"),
"evaluation_level": level,
"deterministic_quality_score": round(deterministic_score, 3),
"quality_grade": grade,
"quality_gate_notes": quality_gates,
"answer_accuracy": round(answer_accuracy, 3),
"source_precision@5": round(mean([float(row["source_precision@5"]) for row in qa_rows]), 3),
"source_recall@5": round(source_recall, 3),
"source_hit@1": round(mean([float(row["source_hit@1"]) for row in qa_rows]), 3),
"source_mrr": round(source_mrr, 3),
"source_ap@5": round(mean([float(row["source_ap@5"]) for row in qa_rows]), 3),
"source_ndcg@5": round(mean([float(row["source_ndcg@5"]) for row in qa_rows]), 3),
"source_duplicate_rate@5": round(mean([float(row["source_duplicate_rate@5"]) for row in qa_rows]), 3),
"context_pruning_precision@5": round(float(adaptive_budget_row.get("source_precision@5", 0.0)), 3),
"context_pruning_recall@5": round(float(adaptive_budget_row.get("source_recall@5", 0.0)), 3),
"context_pruning_token_reduction_pct": round(float(adaptive_budget_row.get("median_context_reduction_pct", 0.0)), 1),
"adaptive_context_target_p50": round(float(adaptive_budget_row.get("median_target_chunks", 0.0)), 1),
"compression_signal_retention": round(float(compression_row.get("answer_signal_retention", 0.0)), 3),
"compression_additional_reduction_pct": round(float(compression_row.get("median_additional_reduction_pct", 0.0)), 1),
"scale_stress_max_chunks": int(largest_scale_row.get("chunks", 0) or 0),
"scale_stress_recall@5": round(float(largest_scale_row.get("source_recall@5", 0.0)), 3) if largest_scale_row else 0.0,
"scale_stress_pruned_recall@5": round(float(largest_scale_row.get("adaptive_pruned_recall@5", 0.0)), 3) if largest_scale_row else 0.0,
"citation_validity": round(citation_validity, 3),
"citation_coverage": round(citation_coverage, 3),
**{key: round(value, 3) for key, value in planner_metrics.items()},
"overview_pass_rate": round(overview_pass, 3),
"abstention_accuracy": round(abstention_accuracy, 3),
"text2sql_pass_rate": round(sql_accuracy, 3),
"hard_mode_pass_rate": round(hard_accuracy, 3),
"profile_benchmark_enabled": bool(profile_rows),
"profile_benchmark_cases": len(profile_rows),
"latency_p50_ms": round(percentile(all_latencies, 0.50), 1),
"latency_p95_ms": round(percentile(all_latencies, 0.95), 1),
"planner_latency_p50_ms": round(percentile(planner_latencies, 0.50), 1),
"planner_latency_p95_ms": round(percentile(planner_latencies, 0.95), 1),
"mean_llm_calls_estimate": round(
mean([float(row.get("llm_calls_estimate", 0)) for row in all_runtime_rows]), 2
),
"focused_context_tokens_before_p50": round(percentile([float(row.get("context_tokens_est_before", 0)) for row in qa_rows], 0.50), 1),
"focused_context_tokens_after_p50": round(percentile([float(row.get("context_tokens_est_after", 0)) for row in qa_rows], 0.50), 1),
"focused_context_pruning_rate": round(mean([float(bool(row.get("context_pruning_used", False))) for row in qa_rows]), 3),
"focused_generation_prompt_tokens_p50": round(percentile([float(row.get("generation_prompt_tokens_est", 0)) for row in qa_rows], 0.50), 1),
"focused_generation_total_tokens_p50": round(percentile([float(row.get("generation_total_tokens_est", 0)) for row in qa_rows], 0.50), 1),
"focused_evidence_utilization_p50": round(percentile([float(row.get("evidence_source_utilization_rate", 0.0)) for row in qa_rows], 0.50), 3),
"focused_evidence_compression_rate": round(mean([float(bool(row.get("evidence_compression_used", False))) for row in qa_rows]), 3),
"focused_evidence_compression_reduction_p50": round(percentile([float(row.get("evidence_compression_reduction_pct", 0.0)) for row in qa_rows], 0.50), 1),
"correction_rate": round(mean([float(row.get("correction_used", False)) for row in all_runtime_rows]), 3),
"runtime_web_use_rate": round(mean([float(row.get("web_used", False)) for row in all_runtime_rows]), 3),
"cache_bypassed": True,
"evaluation_wall_ms": 0.0,
"evaluation_target_rpm": int(pacing_stats["target_rpm"]),
"gemini_requests": int(pacing_stats["gemini_requests"]),
"pacing_sleep_ms": float(pacing_stats["pacing_sleep_ms"]),
"rate_limit_retries": int(pacing_stats["rate_limit_retries"]),
"rate_limit_sleep_ms": float(pacing_stats["rate_limit_sleep_ms"]),
"deep_judge_cases": len(judge_rows),
"scale_stress_error": scale_stress_error,
**{key: round(value, 3) for key, value in judge_summary.items()},
}
node_latency_rows = _node_latency_summary(qa_rows + overview_rows + hard_rows)
profile_summary_rows = _profile_summary(profile_rows)
profile_recommendation = _profile_recommendation(profile_summary_rows)
if profile_recommendation:
summary["profile_recommendation"] = profile_recommendation
if level == "Quick":
readiness_rows = []
summary["release_readiness"] = "NOT RUN"
summary["release_readiness_score"] = None
else:
readiness_rows = _readiness_rows(summary, scale_stress_rows, compression_rows)
readiness_status, readiness_score = _readiness_summary(readiness_rows)
summary["release_readiness"] = readiness_status
summary["release_readiness_score"] = round(readiness_score, 3)
diagnostics = _diagnostics(
summary, ablation_rows, planner_rows, sql_rows, hard_rows, profile_summary_rows, node_latency_rows,
context_budget_rows, compression_rows, scale_stress_rows
)
summary["evaluation_wall_ms"] = round((time.perf_counter() - wall_started) * 1000, 1)
progress(1.0, "Evaluation complete")
return {
"summary": summary,
"diagnostics": diagnostics,
"focused_qa": qa_rows,
"semantic_planner": planner_rows,
"corpus_overviews": overview_rows,
"text2sql": sql_rows,
"abstention": abstention_rows,
"retrieval_ablation": ablation_rows,
"context_budget_ablation": context_budget_rows,
"evidence_compression_ablation": compression_rows,
"scale_stress": scale_stress_rows,
"release_readiness": readiness_rows,
"hard_mode": hard_rows,
"profile_benchmark": profile_rows,
"profile_summary": profile_summary_rows,
"node_latency": node_latency_rows,
"methodology": {
"deterministic": (
"Transparent labels for answer terms, relevant source files, route/task/strategy, web-use policy, "
"citation validity/coverage, abstention and latency. Source metrics deduplicate repeated chunks from "
"the same file before source-level AP/MRR/nDCG are computed."
),
"latency": (
"Evaluation bypasses the response cache and does not mutate chat history, so reported pipeline latency "
"reflects real benchmark execution rather than cached answers. Node-latency summaries subtract deliberate "
"quota pacing proportionally across nodes that issued model calls; raw traces retain wall-clock node time."
),
"deep_judge": (
"Optional Gemini judge for a representative labeled subset of benchmark cases, covering focused QA, "
"NIST, cross-document synthesis and corpus overview. Sampling reduces free-tier request pressure while "
"citation-support and overall scores remain calibrated against deterministic citation validity/coverage. "
"When a compatible Standard report is supplied, Deep reuses that deterministic baseline and only runs "
"the sampled judge layer."
),
"quota_safety": (
f"All Gemini calls in this run share a rolling request pacer targeting {int(pacing_stats['target_rpm'])} RPM. "
"The pacer also accounts for recent interactive requests recorded by this process and surfaced 429s "
"honor provider retry guidance before a bounded retry."
),
"text2sql": (
"Text2SQL routing is evaluated in the semantic-planner suite. The Text2SQL component suite uses one "
"model call per case to generate validated read-only SQL, executes it in DuckDB and checks labeled scalar "
"outputs as typed boolean/numeric/text values when available."
),
"evaluation_cache": (
"Completed reports can be saved by the workspace with corpus/model/benchmark metadata. Saved evaluation "
"history is separate from the RAG response cache and can be reused without rerunning the benchmark."
),
"quality_gates": (
"The letter grade is capped when a critical subsystem is weak, preventing a high weighted average "
"from hiding poor Text2SQL, routing or citation performance."
),
"hard_mode": "Hard-mode cases cover paraphrase, distractors, missing answers, multi-hop comparison, analytical synthesis, structured filtering, local freshness semantics and prompt-injection detection.",
"profile_benchmark": "Optional Fast/Balanced/Agentic comparison uses a small labeled subset because it intentionally spends additional Gemini requests.",
"chunk_ablation": "Retrieval ablation reports source-level metrics plus chunk Hit@1/MRR for cases with explicit chunk-content labels.",
"context_budget": "Focused-query pruning is evaluated as a zero-Gemini ablation across full top-k, a fixed 3-chunk budget, and the adaptive budget chosen from retrieval confidence, score separation and corpus scale.",
"evidence_compression": "Focused evidence compression selects query-relevant sentences after context budgeting and is evaluated by deterministic answer-signal retention plus token reduction; it never spends a Gemini request.",
"scale_stress": "Scale stress reuses existing embedding vectors and clones long-document distractor chunks to exercise the real Qdrant + BM25 path at roughly 1x, 5x and 20x distractor scale without additional Gemini calls.",
"release_readiness": "A transparent readiness checklist applies explicit thresholds to answer quality, grounding, routing, robustness, adaptive-budget recall, compression retention and largest-scale retrieval recall. It does not replace the underlying metrics.",
"benchmark_file": "evals/demo_benchmark.json",
},
}
|