File size: 17,475 Bytes
85b19cf | 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 | """CLI: dataset path, baseline, output dir, dry-run, smoke eval.
Evaluation uses batch LLM judge: 2 calls/session + 2 calls/QA.
Session and QA evaluations run in parallel via ThreadPoolExecutor.
Pipeline results are checkpointed before eval so --eval-only can resume.
"""
from __future__ import annotations
import argparse
import json
import os
from collections.abc import Callable
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import asdict
from pathlib import Path
from typing import Any
try:
from openai import OpenAI
except ImportError:
OpenAI = None # type: ignore[assignment]
from eval_framework.config import EvalConfig
from eval_framework.datasets.domain_a_v2 import (
DomainAV2AcademicBundle,
NormalizedCheckpointQuestion,
load_domain_a_v2_academic,
)
from eval_framework.datasets.schemas import (
MemoryDeltaRecord,
MemorySnapshotRecord,
RetrievalItem,
RetrievalRecord,
)
from eval_framework.evaluators.aggregate import aggregate_metrics
from eval_framework.evaluators.extraction import evaluate_extraction
from eval_framework.evaluators.qa import evaluate_checkpoint_qa
from eval_framework.memory_adapters.base import MemoryAdapter
from eval_framework.openai_compat import (
patch_openai_chat_completions,
rewrite_chat_completion_kwargs,
)
from eval_framework.pipeline.gold_state import GoldMemoryPoint, SessionGoldState
from eval_framework.pipeline.records import PipelineCheckpointQARecord, PipelineSessionRecord
from eval_framework.pipeline.runner import run_domain_a_v2_sample
_CHECKPOINT_SESSIONS = "pipeline_sessions.jsonl"
_CHECKPOINT_QA = "pipeline_qa.jsonl"
# ---------------------------------------------------------------------------
# Checkpoint deserialization: dict -> frozen dataclass
# ---------------------------------------------------------------------------
def _gold_point_from_dict(d: dict[str, Any]) -> GoldMemoryPoint:
return GoldMemoryPoint(
memory_id=d["memory_id"],
memory_content=d["memory_content"],
memory_type=d["memory_type"],
memory_source=d["memory_source"],
is_update=bool(d["is_update"]),
original_memories=tuple(d.get("original_memories") or ()),
importance=float(d.get("importance", 0.0)),
timestamp=d.get("timestamp"),
update_type=d.get("update_type", ""),
)
def _gold_state_from_dict(d: dict[str, Any]) -> SessionGoldState:
return SessionGoldState(
session_id=d["session_id"],
cumulative_gold_memories=tuple(_gold_point_from_dict(g) for g in d["cumulative_gold_memories"]),
session_new_memories=tuple(_gold_point_from_dict(g) for g in d["session_new_memories"]),
session_update_memories=tuple(_gold_point_from_dict(g) for g in d["session_update_memories"]),
session_interference_memories=tuple(_gold_point_from_dict(g) for g in d["session_interference_memories"]),
)
def _snapshot_record_from_dict(d: dict[str, Any]) -> MemorySnapshotRecord:
return MemorySnapshotRecord(
memory_id=d["memory_id"],
text=d["text"],
session_id=d["session_id"],
status=d["status"],
source=d.get("source"),
raw_backend_id=d.get("raw_backend_id"),
raw_backend_type=d.get("raw_backend_type"),
metadata=d.get("metadata") or {},
)
def _delta_record_from_dict(d: dict[str, Any]) -> MemoryDeltaRecord:
return MemoryDeltaRecord(
session_id=d["session_id"],
op=d["op"],
text=d["text"],
linked_previous=tuple(d.get("linked_previous") or ()),
raw_backend_id=d.get("raw_backend_id"),
metadata=d.get("metadata") or {},
)
def _retrieval_item_from_dict(d: dict[str, Any]) -> RetrievalItem:
return RetrievalItem(
rank=int(d["rank"]),
memory_id=d["memory_id"],
text=d["text"],
score=float(d["score"]),
raw_backend_id=d.get("raw_backend_id"),
)
def _retrieval_record_from_dict(d: dict[str, Any]) -> RetrievalRecord:
return RetrievalRecord(
query=d["query"],
top_k=int(d["top_k"]),
items=[_retrieval_item_from_dict(i) for i in d["items"]],
raw_trace=d.get("raw_trace") or {},
)
def _session_record_from_dict(d: dict[str, Any]) -> PipelineSessionRecord:
return PipelineSessionRecord(
sample_id=d["sample_id"],
sample_uuid=d["sample_uuid"],
session_id=d["session_id"],
memory_snapshot=tuple(_snapshot_record_from_dict(s) for s in d["memory_snapshot"]),
memory_delta=tuple(_delta_record_from_dict(dl) for dl in d["memory_delta"]),
gold_state=_gold_state_from_dict(d["gold_state"]),
)
def _qa_record_from_dict(d: dict[str, Any]) -> PipelineCheckpointQARecord:
return PipelineCheckpointQARecord(
sample_id=d["sample_id"],
sample_uuid=d["sample_uuid"],
checkpoint_id=d["checkpoint_id"],
question=d["question"],
gold_answer=d["gold_answer"],
gold_evidence_memory_ids=tuple(d.get("gold_evidence_memory_ids") or ()),
gold_evidence_contents=tuple(d.get("gold_evidence_contents") or ()),
question_type=d["question_type"],
question_type_abbrev=d["question_type_abbrev"],
difficulty=d["difficulty"],
retrieval=_retrieval_record_from_dict(d["retrieval"]),
generated_answer=d["generated_answer"],
cited_memories=tuple(d.get("cited_memories") or ()),
)
def _read_jsonl(path: Path) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
with path.open("r", encoding="utf-8") as fh:
for line in fh:
line = line.strip()
if line:
rows.append(json.loads(line))
return rows
def _load_pipeline_checkpoint(
output_dir: Path,
) -> tuple[list[PipelineSessionRecord], list[PipelineCheckpointQARecord]]:
"""Restore pipeline records from checkpoint JSONL files."""
sess_path = output_dir / _CHECKPOINT_SESSIONS
qa_path = output_dir / _CHECKPOINT_QA
if not sess_path.exists() or not qa_path.exists():
raise SystemExit(
f"Checkpoint files not found in {output_dir}. "
f"Run without --eval-only first to generate them."
)
session_records = [_session_record_from_dict(d) for d in _read_jsonl(sess_path)]
qa_records = [_qa_record_from_dict(d) for d in _read_jsonl(qa_path)]
return session_records, qa_records
def _default_create_adapter(baseline_name: str) -> MemoryAdapter:
from eval_framework.memory_adapters import registry as reg
if baseline_name in reg.MEMGALLERY_NATIVE_REGISTRY:
return reg.MEMGALLERY_NATIVE_REGISTRY[baseline_name]()
if baseline_name in reg.EXTERNAL_ADAPTER_REGISTRY:
return reg.EXTERNAL_ADAPTER_REGISTRY[baseline_name]()
known = sorted(
reg.MEMGALLERY_NATIVE_BASELINES | reg.EXTERNAL_ADAPTER_KEYS
)
raise SystemExit(
f"Unknown baseline {baseline_name!r}. "
f"Expected one of: {', '.join(known)}"
)
def _gold_echo_answer(
q: NormalizedCheckpointQuestion, _retrieval: RetrievalRecord
) -> tuple[str, list[str]]:
return q.gold_answer, []
def _parse_answer_json(raw: str) -> tuple[str, list[str]]:
"""Extract answer and cited_memories from the model's JSON response."""
# Try to parse as JSON first
try:
data = json.loads(raw)
answer = str(data.get("answer", ""))
cited = data.get("cited_memories", [])
if isinstance(cited, list):
return answer, [str(c) for c in cited]
return answer, []
except (json.JSONDecodeError, TypeError):
pass
# Fallback: try to find JSON block in the response
import re
m = re.search(r"\{[\s\S]*\}", raw)
if m:
try:
data = json.loads(m.group())
answer = str(data.get("answer", ""))
cited = data.get("cited_memories", [])
if isinstance(cited, list):
return answer, [str(c) for c in cited]
except (json.JSONDecodeError, TypeError):
pass
# Final fallback: treat entire response as the answer, no citations
return raw.strip(), []
def build_default_answer_fn() -> Callable[
[NormalizedCheckpointQuestion, RetrievalRecord], tuple[str, list[str]]
]:
api_key = os.getenv("OPENAI_API_KEY")
if not api_key or OpenAI is None:
return _gold_echo_answer
client = OpenAI(
api_key=api_key,
base_url=os.getenv("OPENAI_BASE_URL"),
)
model = os.getenv("OPENAI_MODEL") or "gpt-4o"
temperature = float(os.getenv("OPENAI_TEMPERATURE", "0.0"))
max_tokens = int(os.getenv("OPENAI_MAX_TOKENS", "1024"))
def _answer(
q: NormalizedCheckpointQuestion, retrieval: RetrievalRecord
) -> tuple[str, list[str]]:
context_lines = [
f"[{item.rank}] {item.text}" for item in retrieval.items[: retrieval.top_k]
]
context = "\n".join(context_lines) if context_lines else "No retrieved memories."
prompt = (
"Answer the user's question using only the retrieved memories below. "
"If the memories are insufficient, answer exactly: Not mentioned in memory.\n\n"
"You MUST also list the specific memory passages you relied on to produce "
"the answer. Copy the relevant text verbatim from the retrieved memories.\n\n"
f"Question: {q.question}\n\n"
f"Retrieved memories:\n{context}\n\n"
'Respond in JSON:\n'
'{\n'
' "answer": "your concise answer",\n'
' "cited_memories": ["verbatim passage 1", "verbatim passage 2"]\n'
'}\n'
)
request_kwargs = rewrite_chat_completion_kwargs(
{
"model": model,
"messages": [
{
"role": "system",
"content": (
"You answer benchmark questions using only supplied memory context. "
"Be concise and do not invent missing facts. "
"Always respond in the requested JSON format."
),
},
{"role": "user", "content": prompt},
],
"temperature": temperature,
"max_tokens": max_tokens,
}
)
response = client.chat.completions.create(**request_kwargs)
raw = response.choices[0].message.content or ""
return _parse_answer_json(raw)
return _answer
def config_from_namespace(ns: argparse.Namespace) -> EvalConfig:
return EvalConfig(
dataset_path=Path(ns.dataset).expanduser().resolve(),
output_dir=Path(ns.output_dir).expanduser().resolve(),
baseline=str(ns.baseline),
smoke=bool(ns.smoke),
dry_run=bool(ns.dry_run),
)
def _record_to_json_obj(obj: Any) -> dict[str, Any]:
return asdict(obj)
def _write_jsonl(path: Path, rows: list[dict[str, Any]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("w", encoding="utf-8") as fh:
for row in rows:
fh.write(json.dumps(row, ensure_ascii=False) + "\n")
def _write_json(path: Path, payload: dict[str, Any]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(
json.dumps(payload, ensure_ascii=False, indent=2) + "\n",
encoding="utf-8",
)
def run_eval(
config: EvalConfig,
*,
load_domain_bundle: Callable[[Path], DomainAV2AcademicBundle] = load_domain_a_v2_academic,
create_adapter: Callable[[str], MemoryAdapter] | None = None,
answer_fn: Callable | None = None,
max_eval_workers: int = 5,
eval_only: bool = False,
) -> None:
"""Load data, run pipeline (serial) + LLM eval (parallel)."""
patch_openai_chat_completions()
if config.dry_run:
return
out = config.output_dir
out.mkdir(parents=True, exist_ok=True)
if eval_only:
# --- Resume from checkpoint ---
print(f"[Eval-only] Loading pipeline checkpoint from {out}")
session_records, qa_records = _load_pipeline_checkpoint(out)
print(f"[Eval-only] Loaded {len(session_records)} sessions + {len(qa_records)} QA records")
else:
# --- Stage 1: Pipeline (serial — adapter is stateful) ---
adapter_factory = create_adapter or _default_create_adapter
bundle = load_domain_bundle(config.dataset_path)
samples = bundle.samples[:1] if config.smoke else bundle.samples
_answer = answer_fn if answer_fn is not None else build_default_answer_fn()
session_records: list[PipelineSessionRecord] = []
qa_records: list[PipelineCheckpointQARecord] = []
print(f"[Pipeline] Running {len(samples)} sample(s) with baseline={config.baseline}")
for i, sample in enumerate(samples):
print(f" Sample {i + 1}/{len(samples)}: {sample.sample_id}")
adapter = adapter_factory(config.baseline)
sess, qa = run_domain_a_v2_sample(
adapter,
sample,
answer_fn=_answer,
)
session_records.extend(sess)
qa_records.extend(qa)
# --- Save checkpoint ---
_write_jsonl(out / _CHECKPOINT_SESSIONS,
[_record_to_json_obj(r) for r in session_records])
_write_jsonl(out / _CHECKPOINT_QA,
[_record_to_json_obj(r) for r in qa_records])
print(f"[Checkpoint] Saved {len(session_records)} sessions + {len(qa_records)} QA to {out}")
# --- Stage 2: Eval (parallel — each record is self-contained) ---
print(f"[Eval] Evaluating {len(session_records)} sessions + {len(qa_records)} QA with LLM judge (workers={max_eval_workers})...")
session_evals: list[dict[str, object] | None] = [None] * len(session_records)
qa_evals: list[dict[str, object] | None] = [None] * len(qa_records)
with ThreadPoolExecutor(max_workers=max_eval_workers) as pool:
# Submit session evals
session_futures = {}
for idx, srec in enumerate(session_records):
fut = pool.submit(evaluate_extraction, srec)
session_futures[fut] = idx
# Submit QA evals
qa_futures = {}
for idx, qrec in enumerate(qa_records):
fut = pool.submit(evaluate_checkpoint_qa, qrec)
qa_futures[fut] = idx
# Collect session results
done_sessions = 0
for fut in as_completed(session_futures):
idx = session_futures[fut]
try:
session_evals[idx] = fut.result()
except Exception as e:
session_evals[idx] = {"error": str(e)}
done_sessions += 1
if done_sessions % 10 == 0 or done_sessions == len(session_records):
print(f" Sessions: {done_sessions}/{len(session_records)} done")
# Collect QA results
done_qa = 0
for fut in as_completed(qa_futures):
idx = qa_futures[fut]
try:
qa_evals[idx] = fut.result()
except Exception as e:
qa_evals[idx] = {"error": str(e)}
done_qa += 1
if done_qa % 20 == 0 or done_qa == len(qa_records):
print(f" QA: {done_qa}/{len(qa_records)} done")
# --- Stage 3: Aggregate + write ---
agg = aggregate_metrics(
config.baseline,
session_evaluations=[e for e in session_evals if e is not None],
qa_evaluations=[e for e in qa_evals if e is not None],
)
session_rows = []
for srec, s_eval in zip(session_records, session_evals):
row = _record_to_json_obj(srec)
row["eval"] = s_eval
session_rows.append(row)
qa_rows = []
for qrec, q_eval in zip(qa_records, qa_evals):
row = _record_to_json_obj(qrec)
row["eval"] = q_eval
qa_rows.append(row)
_write_jsonl(out / "session_records.jsonl", session_rows)
_write_jsonl(out / "qa_records.jsonl", qa_rows)
_write_json(out / "aggregate_metrics.json", agg)
print(f"\n[Done] Results written to {out}")
print(f" Aggregate: {json.dumps(agg, indent=2)}")
def build_parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(prog="eval_framework")
p.add_argument("--dataset", required=True)
p.add_argument("--baseline", required=True)
p.add_argument("--output-dir", default="eval_framework/results")
p.add_argument("--smoke", action="store_true")
p.add_argument("--dry-run", action="store_true")
p.add_argument("--eval-only", action="store_true",
help="Skip pipeline, load from checkpoint in output-dir.")
p.add_argument("--max-eval-workers", type=int, default=5,
help="Parallel threads for eval stage (default 5).")
return p
def main(argv: list[str] | None = None) -> None:
parser = build_parser()
args = parser.parse_args(argv)
cfg = config_from_namespace(args)
if cfg.dry_run:
print(json.dumps(cfg.to_display_dict(), indent=2))
return
eval_only = bool(args.eval_only)
if not eval_only and not cfg.dataset_path.is_dir():
raise SystemExit(f"Dataset path is not a directory: {cfg.dataset_path}")
run_eval(cfg, max_eval_workers=args.max_eval_workers, eval_only=eval_only)
if __name__ == "__main__":
main()
|