File size: 43,835 Bytes
f440f03 | 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 | """Chat benchmark runner for JSON eval datasets."""
from __future__ import annotations
import asyncio
import json
from collections.abc import Awaitable, Callable
from dataclasses import asdict, dataclass
from datetime import UTC, datetime
from pathlib import Path
from statistics import mean
from time import perf_counter
from typing import Any
import httpx
from maris_core.code.execution_eval import (
CodeExecutionResult,
CodeExecutionSpec,
evaluate_code_response,
)
from maris_core.text.evals import ChatEvalCase, ChatEvalResult, evaluate_chat_case
MEMORY_QUALITY_CATEGORIES = (
"multi_turn_continuity",
"cross_session_recall",
"user_preferences_recall",
"cross_lingual_retrieval",
"stale_memory_rejection",
)
TOOL_USE_TAGS = ("tool", "tools", "tool_use", "grounding", "repo", "browser")
MULTIMODAL_TAGS = ("multimodal", "vision", "image", "video", "audio", "voice", "music")
MULTIMODAL_CATEGORIES = (
"multimodal",
"vision",
"vision_analysis",
"image_generation",
"video_generation",
"voice_conversation",
"audio",
"music_generation",
)
HALLUCINATION_FACTUALITY_THRESHOLD = 0.6
TREND_TRACKED_METRICS = (
"average_overall_score",
"average_latency_ms",
"latency_p95_ms",
"average_tokens_used",
"success_rate",
"reasoning",
"coding",
"safety",
"execution_pass_rate",
"grounding_pass_rate",
"tool_use_pass_rate",
"memory_retrieval_pass_rate",
"multimodal_pass_rate",
"hallucination_rate",
"production_like_pass_rate",
"pairwise_win_rate",
)
@dataclass(frozen=True, slots=True)
class ChatBenchmarkCase:
name: str
message: str
history: tuple[dict[str, str], ...] = ()
vision_context: dict[str, Any] | None = None
profile: str | None = None
persona_id: str | None = None
session_id: str | None = None
expected_terms: tuple[str, ...] = ()
forbidden_terms: tuple[str, ...] = ()
tags: tuple[str, ...] = ()
reference_answer: str | None = None
reference_facts: tuple[str, ...] = ()
expects_code: bool = False
execution_language: str | None = None
execution_test_code: str | None = None
execution_timeout_seconds: float = 8.0
execution_compile_only: bool = False
min_tool_steps: int = 0
min_grounding_sources: int = 0
expected_grounding_terms: tuple[str, ...] = ()
branches: tuple[str, ...] = ()
level: str = "ci"
difficulty: str = "standard"
category: str = "general"
failure_bucket: str = "general"
risk_level: str = "standard"
production_like: bool = False
@dataclass(frozen=True, slots=True)
class ChatBenchmarkResult:
name: str
ok: bool
latency_ms: int
status_code: int | None
response: str
model: str
tokens_used: int
eval: ChatEvalResult | None
execution: CodeExecutionResult | None
grounding_required: bool = False
grounding_ok: bool = True
error: str | None = None
tags: tuple[str, ...] = ()
level: str = "ci"
difficulty: str = "standard"
category: str = "general"
failure_bucket: str = "general"
risk_level: str = "standard"
production_like: bool = False
execution_language: str | None = None
def load_chat_benchmark_dataset(path: str | Path) -> list[ChatBenchmarkCase]:
raw = json.loads(Path(path).read_text(encoding="utf-8"))
entries = raw.get("cases", raw) if isinstance(raw, dict) else raw
if not isinstance(entries, list):
raise ValueError("Benchmark datasetam jābūt JSON masīvam vai objektam ar `cases`.")
cases: list[ChatBenchmarkCase] = []
for entry in entries:
if not isinstance(entry, dict):
raise ValueError("Katram benchmark ierakstam jābūt JSON objektam.")
name = str(entry.get("name", "")).strip()
message = str(entry.get("message", "")).strip()
if not name or not message:
raise ValueError("Benchmark ierakstam obligāti vajag `name` un `message`.")
history = tuple(
item
for item in entry.get("history", [])
if isinstance(item, dict) and item.get("role") and item.get("content")
)
cases.append(
ChatBenchmarkCase(
name=name,
message=message,
history=history,
vision_context=_normalize_optional_mapping(entry.get("vision_context")),
profile=_normalize_optional_text(entry.get("profile")),
persona_id=_normalize_optional_text(entry.get("persona_id")),
session_id=_normalize_optional_text(entry.get("session_id")),
expected_terms=_normalize_text_list(entry.get("expected_terms")),
forbidden_terms=_normalize_text_list(entry.get("forbidden_terms")),
tags=_normalize_text_list(entry.get("tags")),
reference_answer=_normalize_optional_text(entry.get("reference_answer")),
reference_facts=_normalize_text_list(entry.get("reference_facts")),
expects_code=bool(entry.get("expects_code", False)),
execution_language=_normalize_optional_text(entry.get("execution_language")),
execution_test_code=_normalize_optional_text(entry.get("execution_test_code")),
execution_timeout_seconds=float(entry.get("execution_timeout_seconds", 8.0) or 8.0),
execution_compile_only=bool(entry.get("execution_compile_only", False)),
min_tool_steps=max(0, int(entry.get("min_tool_steps", 0) or 0)),
min_grounding_sources=max(0, int(entry.get("min_grounding_sources", 0) or 0)),
expected_grounding_terms=_normalize_text_list(
entry.get("expected_grounding_terms")
),
branches=_normalize_text_list(entry.get("branches")),
level=_normalize_optional_text(entry.get("level")) or "ci",
difficulty=_normalize_optional_text(entry.get("difficulty")) or "standard",
category=_normalize_optional_text(entry.get("category")) or "general",
failure_bucket=_normalize_optional_text(entry.get("failure_bucket")) or "general",
risk_level=_normalize_optional_text(entry.get("risk_level")) or "standard",
production_like=bool(entry.get("production_like", False)),
)
)
return cases
def summarize_chat_benchmark(results: list[ChatBenchmarkResult]) -> dict[str, Any]:
successful = [result for result in results if result.ok]
failed = [result for result in results if not result.ok]
evals = [result.eval for result in successful if result.eval is not None]
category_scores = _group_average_scores(results, key=lambda item: item.category)
failure_bucket_scores = _group_average_scores(results, key=lambda item: item.failure_bucket)
level_scores = _group_average_scores(results, key=lambda item: item.level)
difficulty_scores = _group_average_scores(results, key=lambda item: item.difficulty)
risk_level_scores = _group_average_scores(results, key=lambda item: item.risk_level)
tag_scores = _group_tag_average_scores(results)
tag_pass_rates = _group_tag_pass_rates(results)
execution_supported = [
result for result in results if result.execution is not None and result.execution.available
]
execution_skipped = [
result
for result in results
if result.execution is not None and not result.execution.available
]
execution_passed = [result for result in execution_supported if result.execution.passed]
grounding_cases = [result for result in results if result.grounding_required]
grounding_passed = [
result
for result in grounding_cases
if result.grounding_ok is True and result.error != "grounding requirements not met"
]
execution_language_pass_rates = _group_execution_pass_rates(
execution_supported,
key=lambda item: item.execution_language or "unspecified",
)
execution_language_scores = _group_average_scores(
[result for result in results if result.execution_language],
key=lambda item: item.execution_language or "unspecified",
)
category_execution_pass_rates = _group_execution_pass_rates(
execution_supported,
key=lambda item: item.category,
)
production_like_results = [result for result in results if result.production_like]
production_like_passed = [result for result in production_like_results if result.ok]
memory_results = [
result
for result in results
if result.category in MEMORY_QUALITY_CATEGORIES or "memory" in result.tags
]
memory_passed = [result for result in memory_results if result.ok]
memory_quality_scores = {
key: value for key, value in category_scores.items() if key in MEMORY_QUALITY_CATEGORIES
}
memory_quality_pass_rates = _group_case_pass_rates(
memory_results,
key=lambda item: item.category,
)
tool_use_results = [result for result in results if _is_tool_use_result(result)]
tool_use_passed = [result for result in tool_use_results if result.ok and result.grounding_ok]
multimodal_results = [result for result in results if _is_multimodal_result(result)]
multimodal_passed = [result for result in multimodal_results if result.ok]
hallucination_population = [result for result in results if result.eval is not None]
hallucination_incidents = [
result for result in hallucination_population if _is_hallucination_incident(result)
]
latency_p95_ms = _latency_percentile_ms(results, percentile=95.0)
judge_summary = _summarize_judge_results(evals)
score_manifest = {
"overall": round(mean(item.overall for item in evals), 3) if evals else 0.0,
"reasoning": _mean_eval_metric(evals, "reasoning"),
"factuality": _mean_eval_metric(evals, "factuality"),
"latvian_quality": _mean_eval_metric(evals, "latvian_quality"),
"coding": _mean_eval_metric(evals, "coding"),
"long_context": _mean_eval_metric(evals, "long_context"),
"helpfulness": _mean_eval_metric(evals, "helpfulness"),
"safety": _mean_eval_metric(evals, "safety"),
"tool_use_pass_rate": round(len(tool_use_passed) / len(tool_use_results), 3)
if tool_use_results
else 1.0,
"multimodal_pass_rate": round(len(multimodal_passed) / len(multimodal_results), 3)
if multimodal_results
else 1.0,
"hallucination_rate": round(
len(hallucination_incidents) / len(hallucination_population),
3,
)
if hallucination_population
else 0.0,
"latency_p95_ms": latency_p95_ms,
"execution": round(len(execution_passed) / len(execution_supported), 3)
if execution_supported
else 0.0,
"grounding": round(len(grounding_passed) / len(grounding_cases), 3)
if grounding_cases
else 1.0,
"judge_overall": judge_summary["overall"],
"judge_task_completion": judge_summary["dimension_scores"]["task_completion"],
"judge_instruction_following": judge_summary["dimension_scores"]["instruction_following"],
"judge_grounding": judge_summary["dimension_scores"]["grounding"],
"judge_safety": judge_summary["dimension_scores"]["safety"],
"judge_multi_turn_continuity": judge_summary["dimension_scores"]["multi_turn_continuity"],
"judge_code_quality": judge_summary["dimension_scores"]["code_quality"],
"judge_regression_risk": judge_summary["dimension_scores"]["regression_risk"],
"memory_retrieval_pass_rate": round(len(memory_passed) / len(memory_results), 3)
if memory_results
else 1.0,
}
for category, score in memory_quality_scores.items():
score_manifest[f"memory_{category}"] = score
quality_dimensions = {
"reasoning": {"score": score_manifest["reasoning"], "cases": len(results)},
"coding": {"score": score_manifest["coding"], "cases": len(results)},
"tool_use": {
"cases": len(tool_use_results),
"pass_rate": score_manifest["tool_use_pass_rate"],
},
"memory": {
"cases": len(memory_results),
"pass_rate": score_manifest["memory_retrieval_pass_rate"],
},
"multimodality": {
"cases": len(multimodal_results),
"pass_rate": score_manifest["multimodal_pass_rate"],
},
"latency": {
"average_ms": round(mean(result.latency_ms for result in successful), 1)
if successful
else 0.0,
"p95_ms": latency_p95_ms,
},
"hallucination": {
"cases": len(hallucination_population),
"incident_rate": score_manifest["hallucination_rate"],
},
"safety": {"score": score_manifest["safety"], "cases": len(results)},
}
return {
"total_cases": len(results),
"successful_cases": len(successful),
"failed_cases": len(failed),
"success_rate": round(len(successful) / len(results), 3) if results else 0.0,
"average_latency_ms": round(mean(result.latency_ms for result in successful), 1)
if successful
else 0.0,
"latency_p95_ms": latency_p95_ms,
"average_tokens_used": round(mean(result.tokens_used for result in successful), 1)
if successful
else 0.0,
"average_overall_score": round(mean(item.overall for item in evals), 3) if evals else 0.0,
"execution_cases": len(execution_supported),
"execution_skipped": len(execution_skipped),
"execution_passed": len(execution_passed),
"execution_pass_rate": round(len(execution_passed) / len(execution_supported), 3)
if execution_supported
else 0.0,
"grounding_cases": len(grounding_cases),
"grounding_pass_rate": round(len(grounding_passed) / len(grounding_cases), 3)
if grounding_cases
else 1.0,
"category_scores": category_scores,
"failure_bucket_scores": failure_bucket_scores,
"execution_language_pass_rates": execution_language_pass_rates,
"execution_language_scores": execution_language_scores,
"category_execution_pass_rates": category_execution_pass_rates,
"level_scores": level_scores,
"difficulty_scores": difficulty_scores,
"risk_level_scores": risk_level_scores,
"tag_scores": tag_scores,
"tag_pass_rates": tag_pass_rates,
"memory_retrieval_cases": len(memory_results),
"memory_retrieval_pass_rate": round(len(memory_passed) / len(memory_results), 3)
if memory_results
else 1.0,
"tool_use_cases": len(tool_use_results),
"tool_use_pass_rate": round(len(tool_use_passed) / len(tool_use_results), 3)
if tool_use_results
else 1.0,
"multimodal_cases": len(multimodal_results),
"multimodal_pass_rate": round(len(multimodal_passed) / len(multimodal_results), 3)
if multimodal_results
else 1.0,
"hallucination_eval_cases": len(hallucination_population),
"hallucination_incidents": len(hallucination_incidents),
"hallucination_rate": round(
len(hallucination_incidents) / len(hallucination_population),
3,
)
if hallucination_population
else 0.0,
"memory_quality_scores": memory_quality_scores,
"memory_quality_pass_rates": memory_quality_pass_rates,
"production_like_cases": len(production_like_results),
"production_like_pass_rate": round(
len(production_like_passed) / len(production_like_results), 3
)
if production_like_results
else 1.0,
"quality_dimensions": quality_dimensions,
"judge_summary": judge_summary,
"score_manifest": score_manifest,
"results": [benchmark_result_to_dict(result) for result in results],
}
def build_chat_benchmark_manifest(
results: list[ChatBenchmarkResult],
*,
benchmark_name: str,
branch: str,
model: str,
human_eval_summary: dict[str, Any] | None = None,
) -> dict[str, Any]:
summary = summarize_chat_benchmark(results)
if isinstance(human_eval_summary, dict):
summary["human_eval_summary"] = human_eval_summary
summary["score_manifest"]["pairwise_win_rate"] = round(
float(human_eval_summary.get("pairwise_win_rate", 0.0) or 0.0), 3
)
summary["score_manifest"]["human_eval_confidence"] = round(
float(human_eval_summary.get("average_confidence", 0.0) or 0.0), 3
)
summary["human_eval_cadence"] = _resolve_human_eval_cadence(
benchmark_name=benchmark_name,
branch=branch,
human_eval_summary=human_eval_summary,
)
return {
"benchmark_name": benchmark_name,
"branch": branch,
"model": model,
"generated_at": _utc_timestamp(),
"artifact_type": "chat-benchmark-manifest",
**summary,
}
def build_chat_benchmark_history_artifact(
current_manifest: dict[str, Any],
*,
previous_history: dict[str, Any] | None = None,
max_runs: int = 30,
) -> dict[str, Any]:
existing_runs = _coerce_history_runs(previous_history)
current_run = _build_benchmark_history_run(current_manifest)
baseline = _select_previous_history_baseline(existing_runs, current_run)
runs = list(existing_runs)
if runs and _same_benchmark_identity(runs[-1], current_run):
runs[-1] = current_run
else:
runs.append(current_run)
runs = runs[-max(1, max_runs) :]
regression_report = build_chat_benchmark_regression_report(
current_manifest,
previous_run=baseline,
)
return {
"artifact_type": "chat-benchmark-history",
"benchmark_name": str(current_manifest.get("benchmark_name", "")).strip(),
"branch": str(current_manifest.get("branch", "")).strip(),
"model": str(current_manifest.get("model", "")).strip(),
"latest_generated_at": current_run["generated_at"],
"run_count": len(runs),
"runs": runs,
"trend_summary": _build_benchmark_trend_summary(runs),
"latest_regression_summary": {
"status": regression_report["status"],
"has_baseline": regression_report["has_baseline"],
"has_regressions": regression_report["has_regressions"],
"regression_count": regression_report["regression_count"],
},
}
def build_chat_benchmark_regression_report(
current_manifest: dict[str, Any],
*,
previous_run: dict[str, Any] | None = None,
) -> dict[str, Any]:
current = _extract_benchmark_reference_metrics(current_manifest)
previous = _extract_benchmark_reference_metrics(previous_run)
if previous is None:
return {
"artifact_type": "chat-benchmark-regression-report",
"benchmark_name": current["benchmark_name"],
"branch": current["branch"],
"model": current["model"],
"current_generated_at": current["generated_at"],
"previous_generated_at": None,
"has_baseline": False,
"has_regressions": False,
"regression_count": 0,
"status": "no-baseline",
"score_manifest_deltas": {},
"category_score_deltas": {},
"failure_bucket_score_deltas": {},
"execution_language_pass_rate_deltas": {},
"execution_language_score_deltas": {},
"category_execution_pass_rate_deltas": {},
"memory_quality_score_deltas": {},
"memory_quality_pass_rate_deltas": {},
"regressions": {},
}
score_manifest_deltas = _calculate_metric_deltas(
current["score_manifest"],
previous["score_manifest"],
)
category_score_deltas = _calculate_metric_deltas(
current["category_scores"],
previous["category_scores"],
)
failure_bucket_score_deltas = _calculate_metric_deltas(
current["failure_bucket_scores"],
previous["failure_bucket_scores"],
)
execution_language_pass_rate_deltas = _calculate_metric_deltas(
current["execution_language_pass_rates"],
previous["execution_language_pass_rates"],
)
execution_language_score_deltas = _calculate_metric_deltas(
current["execution_language_scores"],
previous["execution_language_scores"],
)
category_execution_pass_rate_deltas = _calculate_metric_deltas(
current["category_execution_pass_rates"],
previous["category_execution_pass_rates"],
)
memory_quality_score_deltas = _calculate_metric_deltas(
current["memory_quality_scores"],
previous["memory_quality_scores"],
)
memory_quality_pass_rate_deltas = _calculate_metric_deltas(
current["memory_quality_pass_rates"],
previous["memory_quality_pass_rates"],
)
regressions = {
"score_manifest": _filter_negative_deltas(score_manifest_deltas),
"category_scores": _filter_negative_deltas(category_score_deltas),
"failure_bucket_scores": _filter_negative_deltas(failure_bucket_score_deltas),
"execution_language_pass_rates": _filter_negative_deltas(
execution_language_pass_rate_deltas
),
"execution_language_scores": _filter_negative_deltas(execution_language_score_deltas),
"category_execution_pass_rates": _filter_negative_deltas(
category_execution_pass_rate_deltas
),
"memory_quality_scores": _filter_negative_deltas(memory_quality_score_deltas),
"memory_quality_pass_rates": _filter_negative_deltas(memory_quality_pass_rate_deltas),
}
regression_count = sum(len(payload) for payload in regressions.values())
return {
"artifact_type": "chat-benchmark-regression-report",
"benchmark_name": current["benchmark_name"],
"branch": current["branch"],
"model": current["model"],
"current_generated_at": current["generated_at"],
"previous_generated_at": previous["generated_at"],
"has_baseline": True,
"has_regressions": regression_count > 0,
"regression_count": regression_count,
"status": "regression-detected" if regression_count > 0 else "ok",
"score_manifest_deltas": score_manifest_deltas,
"category_score_deltas": category_score_deltas,
"failure_bucket_score_deltas": failure_bucket_score_deltas,
"execution_language_pass_rate_deltas": execution_language_pass_rate_deltas,
"execution_language_score_deltas": execution_language_score_deltas,
"category_execution_pass_rate_deltas": category_execution_pass_rate_deltas,
"memory_quality_score_deltas": memory_quality_score_deltas,
"memory_quality_pass_rate_deltas": memory_quality_pass_rate_deltas,
"regressions": regressions,
}
def select_chat_benchmark_cases(
cases: list[ChatBenchmarkCase],
*,
levels: list[str] | tuple[str, ...],
branch: str | None = None,
) -> list[ChatBenchmarkCase]:
selected = [case for case in cases if case.level in set(levels)]
if not branch:
return selected
normalized_branch = branch.strip().lower()
return [
case
for case in selected
if not case.branches
or normalized_branch in {item.strip().lower() for item in case.branches}
]
async def run_chat_benchmark(
cases: list[ChatBenchmarkCase],
*,
url: str,
concurrency: int = 1,
timeout_seconds: float = 120.0,
transport: httpx.AsyncBaseTransport | None = None,
) -> list[ChatBenchmarkResult]:
async def responder(case: ChatBenchmarkCase) -> dict[str, Any]:
async with httpx.AsyncClient(
timeout=timeout_seconds,
follow_redirects=True,
transport=transport,
) as client:
response = await client.post(
url,
json={
"message": case.message,
"history": list(case.history),
"vision_context": case.vision_context,
"profile": case.profile,
"persona_id": case.persona_id,
"session_id": case.session_id,
},
)
response.raise_for_status()
return response.json()
return await run_chat_benchmark_with_responder(
cases, responder=responder, concurrency=concurrency
)
async def run_chat_benchmark_with_responder(
cases: list[ChatBenchmarkCase],
*,
responder: Callable[[ChatBenchmarkCase], Awaitable[dict[str, Any]]],
concurrency: int = 1,
) -> list[ChatBenchmarkResult]:
semaphore = asyncio.Semaphore(max(1, concurrency))
async def run_case(case: ChatBenchmarkCase) -> ChatBenchmarkResult:
async with semaphore:
return await _execute_case(responder, case)
return await asyncio.gather(*(run_case(case) for case in cases))
def benchmark_result_to_dict(result: ChatBenchmarkResult) -> dict[str, Any]:
payload = asdict(result)
if result.eval is not None:
payload["eval"] = asdict(result.eval)
payload["eval"]["overall"] = result.eval.overall
return payload
def _normalize_optional_text(value: Any) -> str | None:
normalized = str(value or "").strip()
return normalized or None
def _normalize_text_list(value: Any) -> tuple[str, ...]:
if not isinstance(value, list):
return ()
return tuple(str(item).strip() for item in value if str(item).strip())
def _normalize_optional_mapping(value: Any) -> dict[str, Any] | None:
if not isinstance(value, dict):
return None
return dict(value)
async def _execute_case(
responder: Callable[[ChatBenchmarkCase], Awaitable[dict[str, Any]]],
case: ChatBenchmarkCase,
) -> ChatBenchmarkResult:
started_at = perf_counter()
try:
data = await responder(case)
latency_ms = int((perf_counter() - started_at) * 1000)
response_text = str(data.get("response", "")).strip()
execution_result = (
evaluate_code_response(
response_text,
CodeExecutionSpec(
language=case.execution_language,
test_code=case.execution_test_code or "",
timeout_seconds=case.execution_timeout_seconds,
compile_only=case.execution_compile_only,
),
)
if case.execution_language
else None
)
grounding_ok = _grounding_requirements_met(data, case)
eval_result = evaluate_chat_case(
ChatEvalCase(
name=case.name,
prompt=case.message,
response=response_text,
persona_title=str(data.get("persona_title") or "Core Assistant"),
reference_answer=case.reference_answer or "",
reference_facts=case.reference_facts,
expected_terms=case.expected_terms,
forbidden_terms=case.forbidden_terms,
history_turns=len(case.history),
expects_code=case.expects_code,
level=case.level,
difficulty=case.difficulty,
category=case.category,
failure_bucket=case.failure_bucket,
risk_level=case.risk_level,
production_like=case.production_like,
)
)
execution_ok = (
execution_result is None or not execution_result.available or execution_result.passed
)
ok = execution_ok and grounding_ok
return ChatBenchmarkResult(
name=case.name,
ok=ok,
latency_ms=latency_ms,
status_code=int(data.get("status_code", 200) or 200),
response=response_text,
model=str(data.get("model", "")).strip(),
tokens_used=int(data.get("tokens_used", 0) or 0),
eval=eval_result,
execution=execution_result,
grounding_required=_case_requires_grounding(case),
grounding_ok=grounding_ok,
error=_failure_summary(execution_result=execution_result, grounding_ok=grounding_ok),
tags=case.tags,
level=case.level,
difficulty=case.difficulty,
category=case.category,
failure_bucket=case.failure_bucket,
risk_level=case.risk_level,
production_like=case.production_like,
execution_language=case.execution_language,
)
except (httpx.HTTPError, ValueError, TypeError) as exc:
latency_ms = int((perf_counter() - started_at) * 1000)
status_code = getattr(getattr(exc, "response", None), "status_code", None)
return ChatBenchmarkResult(
name=case.name,
ok=False,
latency_ms=latency_ms,
status_code=status_code,
response="",
model="",
tokens_used=0,
eval=None,
execution=None,
grounding_required=_case_requires_grounding(case),
grounding_ok=False,
error=str(exc),
tags=case.tags,
level=case.level,
difficulty=case.difficulty,
category=case.category,
failure_bucket=case.failure_bucket,
risk_level=case.risk_level,
production_like=case.production_like,
execution_language=case.execution_language,
)
def _case_requires_grounding(case: ChatBenchmarkCase) -> bool:
return (
case.min_tool_steps > 0
or case.min_grounding_sources > 0
or bool(case.expected_grounding_terms)
)
def _grounding_requirements_met(data: dict[str, Any], case: ChatBenchmarkCase) -> bool:
if not _case_requires_grounding(case):
return True
tool_trace = data.get("tool_trace")
if not isinstance(tool_trace, dict):
return False
steps = tool_trace.get("steps")
grounding_sources = tool_trace.get("grounding_sources")
if case.min_tool_steps > 0 and (
not isinstance(steps, list) or len(steps) < case.min_tool_steps
):
return False
if case.min_grounding_sources > 0 and (
not isinstance(grounding_sources, list)
or len(grounding_sources) < case.min_grounding_sources
):
return False
if not case.expected_grounding_terms:
return True
haystacks: list[str] = []
if isinstance(grounding_sources, list):
for source in grounding_sources:
if not isinstance(source, dict):
continue
haystacks.extend(
[
str(source.get("label", "")).lower(),
str(source.get("uri", "")).lower(),
str(source.get("snippet", "")).lower(),
]
)
combined = "\n".join(haystacks)
return all(term.lower() in combined for term in case.expected_grounding_terms)
def _failure_summary(
*, execution_result: CodeExecutionResult | None, grounding_ok: bool
) -> str | None:
if execution_result is not None and execution_result.available and not execution_result.passed:
return execution_result.summary
if not grounding_ok:
return "grounding requirements not met"
return None
def _mean_eval_metric(evals: list[ChatEvalResult], field: str) -> float:
if not evals:
return 0.0
return round(mean(float(getattr(item, field, 0.0)) for item in evals), 3)
def _group_average_scores(
results: list[ChatBenchmarkResult],
*,
key: Callable[[ChatBenchmarkResult], str],
) -> dict[str, float]:
buckets: dict[str, list[float]] = {}
for result in results:
if result.eval is None:
continue
bucket = key(result).strip()
if not bucket:
continue
buckets.setdefault(bucket, []).append(result.eval.overall)
return {bucket: round(mean(scores), 3) for bucket, scores in sorted(buckets.items()) if scores}
def _group_case_pass_rates(
results: list[ChatBenchmarkResult],
*,
key: Callable[[ChatBenchmarkResult], str],
) -> dict[str, float]:
buckets: dict[str, list[float]] = {}
for result in results:
bucket = key(result).strip()
if not bucket:
continue
buckets.setdefault(bucket, []).append(1.0 if result.ok else 0.0)
return {bucket: round(mean(scores), 3) for bucket, scores in sorted(buckets.items()) if scores}
def _group_tag_average_scores(results: list[ChatBenchmarkResult]) -> dict[str, float]:
buckets: dict[str, list[float]] = {}
for result in results:
if result.eval is None:
continue
for tag in result.tags:
normalized = str(tag).strip()
if normalized:
buckets.setdefault(normalized, []).append(result.eval.overall)
return {bucket: round(mean(scores), 3) for bucket, scores in sorted(buckets.items()) if scores}
def _group_tag_pass_rates(results: list[ChatBenchmarkResult]) -> dict[str, float]:
buckets: dict[str, list[float]] = {}
for result in results:
for tag in result.tags:
normalized = str(tag).strip()
if normalized:
buckets.setdefault(normalized, []).append(1.0 if result.ok else 0.0)
return {bucket: round(mean(scores), 3) for bucket, scores in sorted(buckets.items()) if scores}
def _group_execution_pass_rates(
results: list[ChatBenchmarkResult],
*,
key: Callable[[ChatBenchmarkResult], str],
) -> dict[str, float]:
buckets: dict[str, list[float]] = {}
for result in results:
if result.execution is None or not result.execution.available:
continue
bucket = key(result).strip()
if not bucket:
continue
buckets.setdefault(bucket, []).append(1.0 if result.execution.passed else 0.0)
return {bucket: round(mean(scores), 3) for bucket, scores in sorted(buckets.items()) if scores}
def _is_tool_use_result(result: ChatBenchmarkResult) -> bool:
if result.grounding_required:
return True
normalized_tags = {str(tag).strip().lower() for tag in result.tags}
return any(tag in normalized_tags for tag in TOOL_USE_TAGS)
def _is_multimodal_result(result: ChatBenchmarkResult) -> bool:
normalized_category = result.category.strip().lower()
if normalized_category in MULTIMODAL_CATEGORIES:
return True
normalized_tags = {str(tag).strip().lower() for tag in result.tags}
return any(tag in normalized_tags for tag in MULTIMODAL_TAGS)
def _is_hallucination_incident(result: ChatBenchmarkResult) -> bool:
if result.eval is None:
return False
if result.eval.factuality < HALLUCINATION_FACTUALITY_THRESHOLD:
return True
judge = result.eval.judge
if judge is None:
return False
return (
result.category in {"grounding", "factuality", "multimodal"} and not judge.grounding.passed
)
def _latency_percentile_ms(
results: list[ChatBenchmarkResult],
*,
percentile: float,
) -> float:
if not results:
return 0.0
latencies = sorted(result.latency_ms for result in results)
index = max(0, min(len(latencies) - 1, round((len(latencies) - 1) * (percentile / 100.0))))
return round(float(latencies[index]), 1)
def _summarize_judge_results(evals: list[ChatEvalResult]) -> dict[str, Any]:
dimension_names = (
"task_completion",
"instruction_following",
"grounding",
"safety",
"multi_turn_continuity",
"code_quality",
"regression_risk",
)
judge_results = [item.judge for item in evals if item.judge is not None]
if not judge_results:
return {
"overall": 0.0,
"pass_rate": 0.0,
"dimension_scores": {name: 0.0 for name in dimension_names},
"failure_reasons": {},
}
failure_reasons: dict[str, int] = {}
for judge in judge_results:
for reason in judge.failure_reasons:
failure_reasons[reason] = failure_reasons.get(reason, 0) + 1
return {
"overall": round(mean(judge.overall for judge in judge_results), 3),
"pass_rate": round(
sum(1 for judge in judge_results if judge.passed) / len(judge_results),
3,
),
"dimension_scores": {
name: round(
mean(float(getattr(judge, name).score) for judge in judge_results),
3,
)
for name in dimension_names
},
"failure_reasons": dict(sorted(failure_reasons.items())),
}
def _utc_timestamp() -> str:
return datetime.now(UTC).replace(microsecond=0).isoformat().replace("+00:00", "Z")
def _normalize_metric_map(value: Any) -> dict[str, float]:
if not isinstance(value, dict):
return {}
normalized: dict[str, float] = {}
for key, item in value.items():
try:
normalized[str(key)] = round(float(item), 3)
except (TypeError, ValueError):
continue
return normalized
def _extract_benchmark_reference_metrics(reference: dict[str, Any] | None) -> dict[str, Any] | None:
if not isinstance(reference, dict):
return None
return {
"benchmark_name": str(reference.get("benchmark_name", "")).strip(),
"branch": str(reference.get("branch", "")).strip(),
"model": str(reference.get("model", "")).strip(),
"generated_at": str(reference.get("generated_at") or _utc_timestamp()),
"score_manifest": _normalize_metric_map(reference.get("score_manifest")),
"category_scores": _normalize_metric_map(reference.get("category_scores")),
"failure_bucket_scores": _normalize_metric_map(reference.get("failure_bucket_scores")),
"tag_scores": _normalize_metric_map(reference.get("tag_scores")),
"tag_pass_rates": _normalize_metric_map(reference.get("tag_pass_rates")),
"execution_language_pass_rates": _normalize_metric_map(
reference.get("execution_language_pass_rates")
),
"execution_language_scores": _normalize_metric_map(
reference.get("execution_language_scores")
),
"category_execution_pass_rates": _normalize_metric_map(
reference.get("category_execution_pass_rates")
),
"memory_quality_scores": _normalize_metric_map(reference.get("memory_quality_scores")),
"memory_quality_pass_rates": _normalize_metric_map(
reference.get("memory_quality_pass_rates")
),
"risk_level_scores": _normalize_metric_map(reference.get("risk_level_scores")),
"average_overall_score": float(reference.get("average_overall_score", 0.0) or 0.0),
"average_latency_ms": float(reference.get("average_latency_ms", 0.0) or 0.0),
"average_tokens_used": float(reference.get("average_tokens_used", 0.0) or 0.0),
"success_rate": float(reference.get("success_rate", 0.0) or 0.0),
"execution_pass_rate": float(reference.get("execution_pass_rate", 0.0) or 0.0),
"grounding_pass_rate": float(reference.get("grounding_pass_rate", 0.0) or 0.0),
"memory_retrieval_pass_rate": float(
reference.get("memory_retrieval_pass_rate", 0.0) or 0.0
),
"production_like_pass_rate": float(reference.get("production_like_pass_rate", 0.0) or 0.0),
"human_eval_cadence": str(reference.get("human_eval_cadence", "") or ""),
"total_cases": int(reference.get("total_cases", 0) or 0),
"failed_cases": int(reference.get("failed_cases", 0) or 0),
}
def _build_benchmark_history_run(manifest: dict[str, Any]) -> dict[str, Any]:
extracted = _extract_benchmark_reference_metrics(manifest)
if extracted is None:
raise ValueError("Benchmark manifest history snapshotam jābūt objektam.")
return extracted
def _coerce_history_runs(history: dict[str, Any] | None) -> list[dict[str, Any]]:
if not isinstance(history, dict):
return []
runs = history.get("runs")
if not isinstance(runs, list):
return []
return [dict(item) for item in runs if isinstance(item, dict)]
def _same_benchmark_identity(left: dict[str, Any], right: dict[str, Any]) -> bool:
return (
str(left.get("benchmark_name", "")).strip() == str(right.get("benchmark_name", "")).strip()
and str(left.get("branch", "")).strip() == str(right.get("branch", "")).strip()
and str(left.get("model", "")).strip() == str(right.get("model", "")).strip()
and str(left.get("generated_at", "")).strip() == str(right.get("generated_at", "")).strip()
)
def _select_previous_history_baseline(
runs: list[dict[str, Any]], current_run: dict[str, Any]
) -> dict[str, Any] | None:
if not runs:
return None
if _same_benchmark_identity(runs[-1], current_run):
return runs[-2] if len(runs) > 1 else None
return runs[-1]
def _calculate_metric_deltas(
current_metrics: dict[str, float],
previous_metrics: dict[str, float],
) -> dict[str, dict[str, float]]:
shared_keys = sorted(set(current_metrics) & set(previous_metrics))
return {
key: {
"current": round(current_metrics[key], 3),
"previous": round(previous_metrics[key], 3),
"delta": round(current_metrics[key] - previous_metrics[key], 3),
}
for key in shared_keys
}
def _filter_negative_deltas(
deltas: dict[str, dict[str, float]],
) -> dict[str, dict[str, float]]:
return {
key: payload
for key, payload in deltas.items()
if float(payload.get("delta", 0.0) or 0.0) < 0.0
}
def _resolve_human_eval_cadence(
*,
benchmark_name: str,
branch: str,
human_eval_summary: dict[str, Any] | None,
) -> str:
if isinstance(human_eval_summary, dict):
cadence = str(human_eval_summary.get("cadence", "") or "").strip()
if cadence:
return cadence
normalized_benchmark = benchmark_name.strip().lower()
normalized_branch = branch.strip().lower()
if "memory" in normalized_benchmark or normalized_branch == "master":
return "weekly + pre-release"
if normalized_branch in {"coder", "planner"}:
return "per release"
return "per release"
def _build_benchmark_trend_summary(runs: list[dict[str, Any]]) -> dict[str, dict[str, Any]]:
if not runs:
return {}
summary: dict[str, dict[str, Any]] = {}
for metric in TREND_TRACKED_METRICS:
recent_values = [_history_metric_value(run, metric) for run in runs]
recent_values = [round(value, 3) for value in recent_values if value is not None]
if not recent_values:
continue
latest = recent_values[-1]
baseline = recent_values[0]
summary[metric] = {
"latest": latest,
"baseline": baseline,
"delta": round(latest - baseline, 3),
"recent_values": recent_values[-5:],
}
return summary
def _history_metric_value(run: dict[str, Any], metric: str) -> float | None:
if metric in run:
try:
return float(run[metric])
except (TypeError, ValueError):
return None
score_manifest = run.get("score_manifest")
if isinstance(score_manifest, dict) and metric in score_manifest:
try:
return float(score_manifest[metric])
except (TypeError, ValueError):
return None
return None
|