@geravant/sinain 1.13.0 → 1.14.0

package/sinain-memory/eval/benchmarks/judges/qa_judge.py

@@ -1,81 +0,0 @@
-"""LLM-as-Judge: QA answer quality evaluator (LongMemEval-compatible, 1-5 scale).
-
-Uses GPT-4o via OpenRouter for comparability with published results.
-"""
-
-from __future__ import annotations
-
-import sys
-from pathlib import Path
-
-# Add sinain-memory to path for common imports
-_koog_dir = str(Path(__file__).resolve().parent.parent.parent.parent)
-if _koog_dir not in sys.path:
-    sys.path.insert(0, _koog_dir)
-
-from common import LLMError, call_llm, extract_json  # noqa: E402
-
-SYSTEM_PROMPT = """\
-You are evaluating whether a predicted answer correctly answers a question.
-The gold (reference) answer is provided.
-
-Score on a scale of 1-5:
-  5: Perfect — captures all key information from the gold answer, no errors
-  4: Mostly correct — minor omissions or imprecision, main point is right
-  3: Partially correct — captures some key points but misses important details
-  2: Related but mostly wrong — touches the topic but answer is largely incorrect
-  1: Completely wrong, contradicts the gold answer, or says "I don't know" when the answer exists
-
-Special cases:
-- If the gold answer indicates abstention is correct (e.g. "I don't know" or "not mentioned"),
-  then a predicted "I don't know" scores 5.
-- Numeric answers within 10% of gold = full credit.
-- Getting the gist right but missing specifics = 3-4 depending on importance.
-
-Respond with ONLY a JSON object: {"score": <1-5>, "reasoning": "brief explanation"}"""
-
-
-def judge_qa(
-    question: str,
-    gold_answer: str,
-    predicted_answer: str,
-    *,
-    condition: str = "",
-    model: str | None = None,
-) -> dict | None:
-    """Score a QA answer. Returns {"score": 1-5, "reasoning": str} or None on failure."""
-    user_parts = [
-        f"## Question\n{question}",
-        f"\n## Gold Answer\n{gold_answer}",
-        f"\n## Predicted Answer\n{predicted_answer}",
-    ]
-    if condition:
-        user_parts.append(f"\n## Context Condition: {condition}")
-
-    try:
-        kwargs: dict = {
-            "system_prompt": SYSTEM_PROMPT,
-            "user_prompt": "\n".join(user_parts),
-            "max_tokens": 200,
-            "json_mode": True,
-        }
-        if model:
-            kwargs["model"] = model
-        else:
-            kwargs["script"] = "meeting_benchmark"
-
-        raw = call_llm(**kwargs)
-        result = extract_json(raw)
-
-        score = result.get("score")
-        reasoning = result.get("reasoning", "")
-
-        if not isinstance(score, (int, float)) or not (1 <= score <= 5):
-            print(f"[warn] qa_judge returned invalid score: {score}", file=sys.stderr)
-            return None
-
-        return {"score": int(score), "reasoning": str(reasoning)[:300]}
-
-    except (ValueError, LLMError, KeyError) as e:
-        print(f"[warn] qa_judge call failed: {e}", file=sys.stderr)
-        return None

package/sinain-memory/eval/benchmarks/longmemeval_adapter.py

@@ -1,177 +0,0 @@
-"""LongMemEval (ICLR 2025) adapter — download + parse into sinain format.
-
-Dataset: https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned
-Paper: https://arxiv.org/abs/2410.10813
-
-Fields per item:
-  question_id, question_type, question, answer, question_date,
-  haystack_session_ids, haystack_dates, haystack_sessions, answer_session_ids
-
-haystack_sessions entries: {"role": "user"/"assistant", "content": "...", "has_answer": bool}
-"""
-
-from __future__ import annotations
-
-import hashlib
-import json
-from datetime import datetime, timedelta, timezone
-from pathlib import Path
-
-from .base_adapter import BenchmarkAdapter, BenchmarkInstance, BenchmarkQuestion
-
-
-def _download_dataset(data_dir: Path) -> Path:
-    """Download LongMemEval from HuggingFace if not cached."""
-    cache_path = data_dir / "longmemeval" / "longmemeval_s_cleaned.json"
-    if cache_path.exists():
-        return cache_path
-
-    cache_path.parent.mkdir(parents=True, exist_ok=True)
-
-    url = "https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_s_cleaned.json"
-    print(f"[longmemeval] downloading from {url} ...")
-
-    # Use curl to avoid macOS Python SSL cert issues
-    import subprocess
-    result = subprocess.run(
-        ["curl", "-fSL", "-o", str(cache_path), url],
-        capture_output=True, text=True, timeout=120,
-    )
-    if result.returncode != 0:
-        raise RuntimeError(f"Download failed: {result.stderr[:200]}")
-    print(f"[longmemeval] saved to {cache_path} ({cache_path.stat().st_size} bytes)")
-    return cache_path
-
-
-def _session_hash(sessions: list[dict]) -> str:
-    """Content hash for a haystack (for grouping questions with shared context)."""
-    raw = json.dumps(sessions, sort_keys=True, ensure_ascii=False)
-    return hashlib.sha256(raw.encode()).hexdigest()[:16]
-
-
-def _sessions_to_feed_items(
-    haystack_sessions: list[list[dict]],
-    haystack_session_ids: list[str],
-    haystack_dates: list[str],
-) -> list[list[dict]]:
-    """Convert LongMemEval haystack into sinain feed item sessions.
-
-    haystack_sessions is a list of sessions, each a list of turn dicts:
-    sessions[i][j] = {"role": "user"/"assistant", "content": "..."}
-
-    Each session becomes a list of feed items with synthesized timestamps.
-    User turns → source: "audio", assistant turns → source: "agent".
-    """
-    result: list[list[dict]] = []
-
-    for i, session_turns in enumerate(haystack_sessions):
-        if not session_turns:
-            continue
-
-        base_ts = haystack_dates[i] if i < len(haystack_dates) else "2025-01-01T10:00:00Z"
-        base_dt = _parse_date(base_ts)
-
-        items = []
-        for j, turn in enumerate(session_turns):
-            ts = (base_dt + timedelta(seconds=30 * j)).isoformat()
-            source = "audio" if turn.get("role") == "user" else "agent"
-            items.append({
-                "source": source,
-                "text": turn.get("content", ""),
-                "ts": ts,
-                "channel": "benchmark",
-            })
-        if items:
-            result.append(items)
-
-    return result
-
-
-def _parse_date(s: str) -> datetime:
-    """Best-effort date parsing."""
-    for fmt in ("%Y-%m-%dT%H:%M:%SZ", "%Y-%m-%dT%H:%M:%S", "%Y-%m-%d", "%m/%d/%Y"):
-        try:
-            return datetime.strptime(s, fmt).replace(tzinfo=timezone.utc)
-        except (ValueError, TypeError):
-            continue
-    return datetime(2025, 1, 1, 10, 0, tzinfo=timezone.utc)
-
-
-class LongMemEvalAdapter(BenchmarkAdapter):
-    """Adapter for LongMemEval (ICLR 2025) benchmark."""
-
-    @property
-    def name(self) -> str:
-        return "longmemeval"
-
-    def load_dataset(self, data_dir: str) -> list[BenchmarkInstance]:
-        """Download and parse LongMemEval, grouping questions by shared haystack."""
-        path = _download_dataset(Path(data_dir))
-        with open(path) as f:
-            raw_items = json.load(f)
-
-        # Group questions by haystack content hash
-        groups: dict[str, dict] = {}
-        for item in raw_items:
-            h = _session_hash(item.get("haystack_sessions", []))
-            if h not in groups:
-                groups[h] = {
-                    "haystack_sessions": item["haystack_sessions"],
-                    "haystack_session_ids": item.get("haystack_session_ids", []),
-                    "haystack_dates": item.get("haystack_dates", []),
-                    "questions": [],
-                }
-            groups[h]["questions"].append(item)
-
-        instances = []
-        for h, group in groups.items():
-            feed_sessions = _sessions_to_feed_items(
-                group["haystack_sessions"],
-                group["haystack_session_ids"],
-                group["haystack_dates"],
-            )
-
-            questions = []
-            for item in group["questions"]:
-                questions.append(BenchmarkQuestion(
-                    id=item["question_id"],
-                    text=item["question"],
-                    gold_answer=str(item["answer"]),
-                    category=item.get("question_type", "unknown"),
-                    evidence_session_ids=item.get("answer_session_ids", []),
-                    metadata={
-                        "question_date": item.get("question_date", ""),
-                    },
-                ))
-
-            instances.append(BenchmarkInstance(
-                id=f"lme-{h}",
-                sessions=feed_sessions,
-                questions=questions,
-                raw_sessions=group["haystack_sessions"],
-                metadata={
-                    "haystack_hash": h,
-                    "num_sessions": len(feed_sessions),
-                    "num_turns": len(group["haystack_sessions"]),
-                },
-            ))
-
-        print(f"[longmemeval] loaded {sum(len(i.questions) for i in instances)} questions "
-              f"across {len(instances)} unique haystacks")
-        return instances
-
-    def format_full_context(self, instance: BenchmarkInstance) -> str:
-        """Render the full conversation history for the baseline condition."""
-        lines = []
-        for session in instance.raw_sessions:
-            if isinstance(session, list):
-                for turn in session:
-                    role = turn.get("role", "unknown").capitalize()
-                    content = turn.get("content", "")
-                    lines.append(f"{role}: {content}")
-                lines.append("---")  # session separator
-            elif isinstance(session, dict):
-                role = session.get("role", "unknown").capitalize()
-                content = session.get("content", "")
-                lines.append(f"{role}: {content}")
-        return "\n\n".join(lines)

package/sinain-memory/eval/benchmarks/meeting_adapter.py

@@ -1,81 +0,0 @@
-"""Meeting Memory adapter — loads QA pairs + ground-truth transcript.
-
-Unlike LongMemEval, this adapter does NOT ingest. The real sinain pipeline
-(start-local.sh) produces the knowledge-graph.db during a live capture run.
-This adapter only loads QA pairs and provides the full-context baseline.
-
-Data layout:
-    eval/benchmarks/data/meeting/
-        <name>.txt          — ground-truth transcript
-        <name>_qa.json      — gold QA pairs
-"""
-
-from __future__ import annotations
-
-import json
-from pathlib import Path
-
-from .base_adapter import BenchmarkAdapter, BenchmarkInstance, BenchmarkQuestion
-
-
-class MeetingMemoryAdapter(BenchmarkAdapter):
-    """Adapter for meeting memory benchmarks (real pipeline capture)."""
-
-    @property
-    def name(self) -> str:
-        return "meeting"
-
-    def load_dataset(self, data_dir: str) -> list[BenchmarkInstance]:
-        """Load meeting QA pairs and transcripts from data/meeting/."""
-        meeting_dir = Path(data_dir) / "meeting"
-        if not meeting_dir.exists():
-            raise FileNotFoundError(f"Meeting data not found: {meeting_dir}")
-
-        instances = []
-        for qa_path in sorted(meeting_dir.glob("*_qa.json")):
-            # Derive transcript path: foo_qa.json → foo.txt
-            stem = qa_path.stem.replace("_qa", "")
-            transcript_path = qa_path.parent / f"{stem}.txt"
-            if not transcript_path.exists():
-                print(f"[meeting] warning: no transcript for {qa_path.name}, skipping")
-                continue
-
-            # Load QA pairs
-            with open(qa_path) as f:
-                raw_questions = json.load(f)
-
-            questions = []
-            for item in raw_questions:
-                questions.append(BenchmarkQuestion(
-                    id=item["id"],
-                    text=item["question"],
-                    gold_answer=item["gold_answer"],
-                    category=item.get("category", "unknown"),
-                    evidence_session_ids=item.get("evidence_timestamps", []),
-                    metadata={
-                        "evidence_timestamps": item.get("evidence_timestamps", []),
-                    },
-                ))
-
-            # Load transcript
-            transcript_text = transcript_path.read_text(encoding="utf-8")
-
-            instances.append(BenchmarkInstance(
-                id=f"meeting-{stem}",
-                sessions=[],  # Not used — real pipeline does ingestion
-                questions=questions,
-                raw_sessions=[],
-                metadata={
-                    "transcript_path": str(transcript_path),
-                    "raw_transcript": transcript_text,
-                    "stem": stem,
-                },
-            ))
-
-        total_q = sum(len(i.questions) for i in instances)
-        print(f"[meeting] loaded {total_q} questions across {len(instances)} meetings")
-        return instances
-
-    def format_full_context(self, instance: BenchmarkInstance) -> str:
-        """Return the raw transcript verbatim for the full-context baseline."""
-        return instance.metadata.get("raw_transcript", "")

package/sinain-memory/eval/benchmarks/meeting_runner.py

@@ -1,230 +0,0 @@
-#!/usr/bin/env python3
-"""Meeting Memory Benchmark runner — evaluates a captured knowledge-graph.db.
-
-Standalone script: does NOT modify or import runner.py.
-Reuses shared infrastructure: query, judge, evaluate, report.
-
-Usage:
-    python3 eval/benchmarks/meeting_runner.py \
-        --db /tmp/sinain-bench-XXXX/knowledge-graph.db \
-        --conditions sinain-memory,full-context
-
-    # Quick test (3 questions):
-    python3 eval/benchmarks/meeting_runner.py \
-        --db /tmp/sinain-bench-XXXX/knowledge-graph.db \
-        --subset 3
-"""
-
-from __future__ import annotations
-
-import argparse
-import json
-import sys
-from pathlib import Path
-
-# Add sinain-memory to path
-_koog_dir = str(Path(__file__).resolve().parent.parent.parent)
-if _koog_dir not in sys.path:
-    sys.path.insert(0, _koog_dir)
-
-from eval.benchmarks.config import DATA_DIR, RESULTS_DIR, QA_MODEL, JUDGE_MODEL
-from eval.benchmarks.meeting_adapter import MeetingMemoryAdapter
-from eval.benchmarks.query import answer_question, _get_retrieved_facts, compute_content_recall
-from eval.benchmarks.evaluate import token_f1, aggregate_results
-from eval.benchmarks.judges.qa_judge import judge_qa
-from eval.benchmarks.report import generate_markdown, generate_json
-
-
-def run_meeting_benchmark(
-    db_path: str,
-    conditions: list[str],
-    *,
-    subset: int | None = None,
-    meeting_filter: str | None = None,
-    qa_model: str = QA_MODEL,
-    judge_model: str = JUDGE_MODEL,
-    output_dir: Path = RESULTS_DIR,
-    data_dir: Path = DATA_DIR,
-    resume: bool = False,
-) -> tuple[dict, list[dict]]:
-    """Run meeting benchmark against a captured knowledge-graph.db."""
-
-    adapter = MeetingMemoryAdapter()
-    instances = adapter.load_dataset(str(data_dir))
-
-    if meeting_filter:
-        instances = [i for i in instances if meeting_filter in i.id]
-        print(f"[meeting] filtered to {len(instances)} meeting(s) matching '{meeting_filter}'")
-
-    if not instances:
-        print("[meeting] no meeting data found — check data/meeting/ directory")
-        return {"error": "no data"}, []
-
-    # Validate DB exists for sinain-memory condition
-    if "sinain-memory" in conditions:
-        if not db_path or not Path(db_path).exists():
-            print(f"[meeting] ERROR: --db path does not exist: {db_path}")
-            print("  Run the capture pipeline first (see plan Part 1)")
-            sys.exit(1)
-
-    print(f"\n{'='*60}")
-    print(f"  Meeting Memory Benchmark")
-    print(f"  Conditions: {', '.join(conditions)}")
-    print(f"  DB: {db_path or '(none)'}")
-    print(f"  QA model: {qa_model}")
-    print(f"  Judge model: {judge_model}")
-    print(f"{'='*60}\n")
-
-    # Flatten questions
-    all_questions = []
-    for inst in instances:
-        for q in inst.questions:
-            all_questions.append((inst, q))
-
-    if subset:
-        all_questions = all_questions[:subset]
-
-    total = len(all_questions)
-    print(f"[meeting] evaluating {total} questions\n")
-
-    # Resume support
-    resume_path = output_dir / "meeting_progress.jsonl"
-    completed: dict[str, dict] = {}
-    if resume and resume_path.exists():
-        for line in resume_path.read_text().strip().split("\n"):
-            if line:
-                entry = json.loads(line)
-                completed[entry["id"]] = entry
-
-    output_dir.mkdir(parents=True, exist_ok=True)
-    details: list[dict] = []
-
-    for idx, (inst, question) in enumerate(all_questions):
-        qid = question.id
-
-        if qid in completed:
-            details.append(completed[qid])
-            continue
-
-        print(f"[{idx+1}/{total}] {qid} [{question.category}]")
-        print(f"  Q: {question.text[:80]}...")
-
-        full_context = adapter.format_full_context(inst)
-
-        # Retrieval metrics
-        retrieval = {}
-        if db_path and "sinain-memory" in conditions:
-            retrieved_facts = _get_retrieved_facts(db_path, question.text)
-            retrieval = compute_content_recall(retrieved_facts, question.gold_answer)
-
-        # Generate answers per condition
-        answers = {}
-        for cond in conditions:
-            if cond == "sinain-memory" and not db_path:
-                answers[cond] = {"text": "(no DB)", "score": 1, "f1": 0.0}
-                continue
-
-            print(f"  [{cond}] generating answer...")
-            answer_text = answer_question(
-                question, cond,
-                db_path=db_path,
-                full_context=full_context,
-                model=qa_model,
-            )
-
-            f1 = token_f1(answer_text, question.gold_answer)
-
-            judge_result = judge_qa(
-                question.text, question.gold_answer, answer_text,
-                condition=cond, model=judge_model,
-            )
-            score = judge_result["score"] if judge_result else None
-            reasoning = judge_result["reasoning"] if judge_result else None
-
-            answers[cond] = {
-                "text": answer_text[:500],
-                "score": score,
-                "f1": round(f1, 4),
-                "reasoning": reasoning,
-            }
-            print(f"    score={score}/5 f1={f1:.2f}")
-
-        entry = {
-            "id": qid,
-            "question": question.text,
-            "gold_answer": question.gold_answer,
-            "category": question.category,
-            "retrieval": retrieval,
-            "answers": answers,
-        }
-        details.append(entry)
-
-        # Save progress
-        with open(resume_path, "a") as f:
-            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
-
-    summary = aggregate_results(details)
-    return summary, details
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser(description="Meeting Memory Benchmark")
-    parser.add_argument("--db", required=False, default=None,
-                        help="Path to knowledge-graph.db from capture run")
-    parser.add_argument("--conditions", default="sinain-memory,full-context",
-                        help="Comma-separated conditions (sinain-memory, full-context)")
-    parser.add_argument("--subset", type=int, default=None,
-                        help="Run only first N questions")
-    parser.add_argument("--meeting", default=None,
-                        help="Filter to specific meeting by stem (e.g. al-futaim-prep-5min)")
-    parser.add_argument("--qa-model", default=QA_MODEL)
-    parser.add_argument("--judge-model", default=JUDGE_MODEL)
-    parser.add_argument("--output-dir", type=Path, default=RESULTS_DIR)
-    parser.add_argument("--data-dir", type=Path, default=DATA_DIR)
-    parser.add_argument("--format", default="json,markdown")
-    parser.add_argument("--resume", action="store_true")
-    args = parser.parse_args()
-
-    conditions = [c.strip() for c in args.conditions.split(",")]
-    formats = [f.strip() for f in args.format.split(",")]
-
-    summary, details = run_meeting_benchmark(
-        args.db, conditions,
-        subset=args.subset,
-        meeting_filter=args.meeting,
-        qa_model=args.qa_model,
-        judge_model=args.judge_model,
-        output_dir=args.output_dir,
-        data_dir=args.data_dir,
-        resume=args.resume,
-    )
-
-    # Write outputs
-    args.output_dir.mkdir(parents=True, exist_ok=True)
-
-    if "json" in formats:
-        json_path = args.output_dir / "meeting_results.json"
-        json_path.write_text(generate_json("meeting", summary, details))
-        print(f"\n[output] JSON: {json_path}")
-
-    if "markdown" in formats:
-        md_path = args.output_dir / "meeting_results.md"
-        md_path.write_text(generate_markdown("meeting", summary, details))
-        print(f"[output] Markdown: {md_path}")
-
-    # Print summary
-    print(f"\n{'='*60}")
-    print(f"  Meeting Memory Benchmark — Summary")
-    print(f"{'='*60}")
-    ipr = summary.get("ipr")
-    if ipr:
-        print(f"  IPR: {ipr:.1%}")
-    for cond, data in summary.get("conditions", {}).items():
-        print(f"  {cond}: {data['mean_score']:.2f}/5 (n={data['n']})")
-    for k, v in summary.get("retrieval", {}).items():
-        print(f"  {k}: {v:.1%}")
-    print()
-
-
-if __name__ == "__main__":
-    main()