@geravant/sinain 1.13.0 → 1.14.0

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

@@ -1,283 +0,0 @@
-#!/usr/bin/env python3
-"""Benchmark runner — evaluates sinain's knowledge graph against published benchmarks.
-
-Usage:
-    python3 eval/benchmarks/runner.py --benchmarks longmemeval --subset 5
-    python3 eval/benchmarks/runner.py --benchmarks longmemeval --conditions sinain-memory,full-context
-    python3 eval/benchmarks/runner.py --benchmarks longmemeval --format markdown --resume
-"""
-
-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.base_adapter import BenchmarkAdapter, BenchmarkInstance
-from eval.benchmarks.longmemeval_adapter import LongMemEvalAdapter
-from eval.benchmarks.ingest import ingest_instance, get_knowledge_doc
-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 _get_adapter(name: str) -> BenchmarkAdapter:
-    if name == "longmemeval":
-        return LongMemEvalAdapter()
-    raise ValueError(f"Unknown benchmark: {name}. Available: longmemeval")
-
-
-def _load_resume(resume_path: Path) -> dict[str, dict]:
-    """Load previously computed results for resume support."""
-    results = {}
-    if resume_path.exists():
-        for line in resume_path.read_text().strip().split("\n"):
-            if line:
-                entry = json.loads(line)
-                results[entry["id"]] = entry
-    return results
-
-
-def run_benchmark(
-    benchmark_name: str,
-    conditions: list[str],
-    *,
-    subset: int | None = None,
-    qa_model: str = QA_MODEL,
-    judge_model: str = JUDGE_MODEL,
-    output_dir: Path = RESULTS_DIR,
-    cache_dir: Path = DATA_DIR,
-    resume: bool = False,
-    skip_llm: bool = False,
-    stratified: bool = False,
-) -> tuple[dict, list[dict]]:
-    """Run a benchmark end-to-end. Returns (summary, details)."""
-
-    adapter = _get_adapter(benchmark_name)
-
-    # Load dataset
-    print(f"\n{'='*60}")
-    print(f"  Benchmark: {benchmark_name}")
-    print(f"  Conditions: {', '.join(conditions)}")
-    print(f"  QA model: {qa_model}")
-    print(f"  Judge model: {judge_model}")
-    print(f"{'='*60}\n")
-
-    instances = adapter.load_dataset(str(cache_dir))
-
-    # Flatten questions
-    all_questions = []
-    for inst in instances:
-        for q in inst.questions:
-            all_questions.append((inst, q))
-
-    if subset:
-        if stratified:
-            # Take equal samples from each question category
-            from collections import defaultdict
-            by_cat: dict[str, list] = defaultdict(list)
-            for pair in all_questions:
-                by_cat[pair[1].category].append(pair)
-            per_cat = max(1, subset // len(by_cat))
-            sampled = []
-            for cat in sorted(by_cat):
-                sampled.extend(by_cat[cat][:per_cat])
-            all_questions = sampled[:subset]
-        else:
-            all_questions = all_questions[:subset]
-
-    total = len(all_questions)
-    print(f"[runner] evaluating {total} questions\n")
-
-    # Resume support
-    resume_path = output_dir / f"{benchmark_name}_progress.jsonl"
-    completed = _load_resume(resume_path) if resume else {}
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    # Track ingested instances
-    instance_dbs: dict[str, Path | None] = {}
-    instance_docs: dict[str, str] = {}
-
-    details: list[dict] = []
-
-    for idx, (inst, question) in enumerate(all_questions):
-        qid = question.id
-
-        # Skip if already done (with all conditions scored)
-        if qid in completed:
-            prev = completed[qid]
-            all_scored = all(
-                prev.get("answers", {}).get(c, {}).get("score") is not None
-                for c in conditions
-            )
-            if all_scored:
-                details.append(prev)
-                continue
-            # Otherwise re-run this question (previous attempt had failures)
-
-        print(f"[{idx+1}/{total}] {qid} [{question.category}]")
-
-        # Ingest instance if not done yet
-        if inst.id not in instance_dbs:
-            if "sinain-memory" in conditions or "knowledge-doc" in conditions:
-                print(f"  ingesting {inst.id} ({len(inst.sessions)} sessions)...")
-                instance_dbs[inst.id] = ingest_instance(inst, cache_dir / benchmark_name)
-                db = instance_dbs[inst.id]
-                if db:
-                    instance_docs[inst.id] = get_knowledge_doc(db)
-                    print(f"  -> ingested ({db.stat().st_size} bytes)")
-                else:
-                    instance_docs[inst.id] = "(ingestion failed)"
-                    print(f"  -> ingestion failed")
-            else:
-                instance_dbs[inst.id] = None
-                instance_docs[inst.id] = ""
-
-        db_path = instance_dbs.get(inst.id)
-        knowledge_doc = instance_docs.get(inst.id, "")
-        full_context = adapter.format_full_context(inst)
-
-        # Retrieval metrics (content-based: do retrieved facts contain the answer?)
-        retrieval = {}
-        if db_path and "sinain-memory" in conditions:
-            retrieved_facts = _get_retrieved_facts(str(db_path), question.text)
-            retrieval = compute_content_recall(
-                retrieved_facts, question.gold_answer,
-            )
-
-        # Generate answers per condition
-        answers = {}
-        for cond in conditions:
-            if skip_llm:
-                answers[cond] = {"text": "(skipped)", "score": None, "f1": None}
-                continue
-
-            # Skip sinain-memory/knowledge-doc if ingestion failed
-            if cond in ("sinain-memory", "knowledge-doc") and not db_path:
-                answers[cond] = {"text": "(ingestion failed)", "score": 1, "f1": 0.0, "reasoning": "ingestion failed"}
-                print(f"  [{cond}] skipped (ingestion failed)")
-                continue
-
-            print(f"  [{cond}] generating answer...")
-            answer_text = answer_question(
-                question, cond,
-                db_path=str(db_path) if db_path else None,
-                full_context=full_context,
-                knowledge_doc=knowledge_doc,
-                model=qa_model,
-            )
-
-            # Score
-            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 incrementally
-        with open(resume_path, "a") as f:
-            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
-
-    # Aggregate
-    summary = aggregate_results(details)
-    return summary, details
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser(description="Sinain Knowledge Graph Benchmark Runner")
-    parser.add_argument("--benchmarks", default="longmemeval",
-                        help="Comma-separated benchmark names (longmemeval, locomo)")
-    parser.add_argument("--conditions", default="sinain-memory,full-context,knowledge-doc",
-                        help="Comma-separated conditions to evaluate")
-    parser.add_argument("--subset", type=int, default=None,
-                        help="Run only first N questions (for dev iteration)")
-    parser.add_argument("--qa-model", default=QA_MODEL, help="Model for QA generation")
-    parser.add_argument("--judge-model", default=JUDGE_MODEL, help="Model for QA judging")
-    parser.add_argument("--output-dir", type=Path, default=RESULTS_DIR)
-    parser.add_argument("--cache-dir", type=Path, default=DATA_DIR)
-    parser.add_argument("--format", default="json,markdown",
-                        help="Output formats (json, markdown)")
-    parser.add_argument("--resume", action="store_true", help="Resume from partial results")
-    parser.add_argument("--skip-llm", action="store_true",
-                        help="Skip LLM calls (retrieval + mechanical metrics only)")
-    parser.add_argument("--stratified", action="store_true",
-                        help="Sample equally from each question category (with --subset)")
-    args = parser.parse_args()
-
-    conditions = [c.strip() for c in args.conditions.split(",")]
-    formats = [f.strip() for f in args.format.split(",")]
-
-    for bench_name in args.benchmarks.split(","):
-        bench_name = bench_name.strip()
-        summary, details = run_benchmark(
-            bench_name, conditions,
-            subset=args.subset,
-            qa_model=args.qa_model,
-            judge_model=args.judge_model,
-            output_dir=args.output_dir,
-            cache_dir=args.cache_dir,
-            resume=args.resume,
-            skip_llm=args.skip_llm,
-            stratified=args.stratified,
-        )
-
-        # Write outputs
-        args.output_dir.mkdir(parents=True, exist_ok=True)
-
-        if "json" in formats:
-            json_path = args.output_dir / f"{bench_name}_results.json"
-            json_path.write_text(generate_json(bench_name, summary, details))
-            print(f"\n[output] JSON: {json_path}")
-
-        if "markdown" in formats:
-            md_path = args.output_dir / f"{bench_name}_results.md"
-            md_path.write_text(generate_markdown(bench_name, summary, details))
-            print(f"[output] Markdown: {md_path}")
-
-        # Print summary
-        print(f"\n{'='*60}")
-        print(f"  {bench_name} — 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()

package/sinain-memory/eval/judges/base_judge.py

@@ -1,61 +0,0 @@
-"""Shared infrastructure for LLM-as-Judge evaluators.
-
-Provides ``run_judge()`` which calls the LLM with a rubric prompt and
-extracts a ``{"score": 1-4, "reasoning": "..."}`` response.
-"""
-
-from __future__ import annotations
-
-import json
-import sys
-from pathlib import Path
-
-# Add parent dirs so ``common`` is importable when running from anywhere.
-_koog_dir = str(Path(__file__).resolve().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
-
-
-def run_judge(
-    system_prompt: str,
-    user_prompt: str,
-    *,
-    model: str | None = None,
-    max_tokens: int = 200,
-    timeout: int = 30,
-) -> dict | None:
-    """Call LLM with a judge prompt and return ``{"score": int, "reasoning": str}`` or None.
-
-    *model* defaults to the ``eval.judges.model`` setting resolved externally.
-    When None, falls back to ``common.call_llm`` defaults (which reads koog-config).
-    """
-    try:
-        kwargs: dict = {
-            "system_prompt": system_prompt,
-            "user_prompt": user_prompt,
-            "max_tokens": max_tokens,
-            "json_mode": True,
-        }
-        # Use script-based config resolution if no explicit model
-        if model:
-            kwargs["model"] = model
-        else:
-            kwargs["script"] = "tick_evaluator"
-
-        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 <= 4):
-            print(f"[warn] 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] judge call failed: {e}", file=sys.stderr)
-        return None

package/sinain-memory/eval/judges/curation_judge.py

@@ -1,46 +0,0 @@
-"""LLM-as-Judge: Playbook curation quality evaluator."""
-
-from __future__ import annotations
-
-from .base_judge import run_judge
-
-SYSTEM_PROMPT = """\
-You are an evaluator scoring the quality of playbook curation changes.
-
-The curator follows a directive and three laws:
-  Law 1: Don't remove error-prevention patterns
-  Law 2: Preserve high-scoring approaches
-  Law 3: Then evolve
-
-Rate the curation on a 1-4 scale:
-  4: Changes perfectly match directive + evidence, three laws respected
-  3: Good changes, minor alignment issues with directive
-  2: Changes misaligned with directive or weak evidence
-  1: Destructive changes, violated three laws, or ignored directive entirely
-
-Respond with ONLY a JSON object: {"score": <1-4>, "reasoning": "brief explanation"}"""
-
-
-def judge_curation(
-    curator_result: dict,
-    directive: str,
-    playbook_before: str = "",
-    **kwargs,
-) -> dict | None:
-    """Evaluate playbook curation quality. Returns {"score": 1-4, "reasoning": str} or None."""
-    changes = curator_result.get("changes", {})
-    stale_actions = curator_result.get("staleItemActions", [])
-    lines = curator_result.get("playbookLines", "?")
-
-    parts = [
-        f"## Curate Directive\n{directive}",
-        f"\n## Changes Made\nAdded: {changes.get('added', [])}\nPruned: {changes.get('pruned', [])}\nPromoted: {changes.get('promoted', [])}",
-        f"\n## Stale Item Actions\n{stale_actions}",
-        f"\n## Playbook Lines After: {lines}",
-    ]
-
-    if playbook_before:
-        # Truncate to keep prompt manageable
-        parts.append(f"\n## Playbook Before (excerpt)\n{playbook_before[:1500]}")
-
-    return run_judge(SYSTEM_PROMPT, "\n".join(parts), **kwargs)

package/sinain-memory/eval/judges/insight_judge.py

@@ -1,48 +0,0 @@
-"""LLM-as-Judge: Insight synthesis quality evaluator."""
-
-from __future__ import annotations
-
-from .base_judge import run_judge
-
-SYSTEM_PROMPT = """\
-You are an evaluator scoring the quality of an insight synthesizer's output.
-
-The synthesizer produces two parts:
-  - Suggestion: actionable recommendation grounded in playbook/data
-  - Insight: surprising cross-domain connection from accumulated observations
-
-Rate the output on a 1-4 scale:
-  4: Suggestion is actionable with specific reference, insight connects 2+ distinct observations
-  3: One component is excellent, the other adequate
-  2: Generic suggestion or obvious insight
-  1: Hallucinated content, not grounded in playbook/logs
-
-If the output was skipped, rate the skip decision:
-  4: Skip is well-justified with specific references to what was checked
-  3: Skip is reasonable
-  2: Should not have skipped — there was material to work with
-  1: Skip reason is generic/lazy
-
-Respond with ONLY a JSON object: {"score": <1-4>, "reasoning": "brief explanation"}"""
-
-
-def judge_insight(
-    synth_result: dict,
-    playbook_excerpt: str = "",
-    **kwargs,
-) -> dict | None:
-    """Evaluate insight synthesis quality. Returns {"score": 1-4, "reasoning": str} or None."""
-    skipped = synth_result.get("skip", False)
-
-    parts = []
-    if skipped:
-        parts.append(f"## Status: SKIPPED\nReason: {synth_result.get('skipReason', 'none given')}")
-    else:
-        parts.append(f"## Suggestion\n{synth_result.get('suggestion', '')}")
-        parts.append(f"\n## Insight\n{synth_result.get('insight', '')}")
-        parts.append(f"\n## Total Chars: {synth_result.get('totalChars', '?')}")
-
-    if playbook_excerpt:
-        parts.append(f"\n## Playbook Context (excerpt)\n{playbook_excerpt[:1000]}")
-
-    return run_judge(SYSTEM_PROMPT, "\n".join(parts), **kwargs)

package/sinain-memory/eval/judges/mining_judge.py

@@ -1,42 +0,0 @@
-"""LLM-as-Judge: Memory mining quality evaluator."""
-
-from __future__ import annotations
-
-from .base_judge import run_judge
-
-SYSTEM_PROMPT = """\
-You are an evaluator scoring the quality of a memory mining agent's findings.
-
-The miner reads daily memory files and extracts patterns, preferences, and insights
-that should be added to the evolving playbook.
-
-Rate the mining output on a 1-4 scale:
-  4: Found non-obvious cross-day patterns, all grounded in source files
-  3: Valid patterns found, properly grounded in provided daily files
-  2: Only surface-level observations from source files
-  1: Hallucinated patterns not present in provided daily files
-
-Respond with ONLY a JSON object: {"score": <1-4>, "reasoning": "brief explanation"}"""
-
-
-def judge_mining(
-    miner_result: dict,
-    mined_file_excerpts: dict[str, str] | None = None,
-    **kwargs,
-) -> dict | None:
-    """Evaluate memory mining quality. Returns {"score": 1-4, "reasoning": str} or None."""
-    parts = [
-        f"## Findings\n{miner_result.get('findings', '')}",
-        f"\n## New Patterns\n{miner_result.get('newPatterns', [])}",
-        f"\n## Contradictions\n{miner_result.get('contradictions', [])}",
-        f"\n## Preferences\n{miner_result.get('preferences', [])}",
-        f"\n## Mined Sources\n{miner_result.get('minedSources', [])}",
-    ]
-
-    if mined_file_excerpts:
-        for name, content in mined_file_excerpts.items():
-            # Truncate large files
-            excerpt = content[:1500] if len(content) > 1500 else content
-            parts.append(f"\n## Source File: {name}\n{excerpt}")
-
-    return run_judge(SYSTEM_PROMPT, "\n".join(parts), **kwargs)

package/sinain-memory/eval/judges/signal_judge.py

@@ -1,45 +0,0 @@
-"""LLM-as-Judge: Signal detection quality evaluator."""
-
-from __future__ import annotations
-
-from .base_judge import run_judge
-
-SYSTEM_PROMPT = """\
-You are an evaluator scoring the quality of a signal detection system.
-
-Rate the signal analysis on a 1-4 scale:
-  4: All real signals detected, action is highly relevant and specific
-  3: Key signals detected, action is reasonable
-  2: Missed important signals or action is vague
-  1: Hallucinated signals or inappropriate action
-
-Respond with ONLY a JSON object: {"score": <1-4>, "reasoning": "brief explanation"}"""
-
-
-def judge_signal(
-    signal_result: dict,
-    session_summary: str,
-    recent_logs: list[dict] | None = None,
-    **kwargs,
-) -> dict | None:
-    """Evaluate signal detection quality. Returns {"score": 1-4, "reasoning": str} or None."""
-    parts = [f"## Session Summary\n{session_summary}"]
-
-    signals = signal_result.get("signals", [])
-    action = signal_result.get("recommendedAction")
-    idle = signal_result.get("idle", False)
-
-    parts.append(f"\n## Detected Signals\n{signals}")
-    parts.append(f"\n## Recommended Action\n{action}")
-    parts.append(f"\n## Idle: {idle}")
-
-    if recent_logs:
-        recent_actions = []
-        for log in recent_logs[:3]:
-            for a in log.get("actionsConsidered", []):
-                if a.get("chosen"):
-                    recent_actions.append(a)
-        if recent_actions:
-            parts.append(f"\n## Recent Actions (should not repeat)\n{recent_actions}")
-
-    return run_judge(SYSTEM_PROMPT, "\n".join(parts), **kwargs)

package/sinain-memory/eval/retrieval_benchmark.jsonl

@@ -1,12 +0,0 @@
-{"query": "OCR pipeline stalls on macOS 14", "expected_entities": ["fact:ocr-backpressure", "fact:sck-capture"], "category": "error-resolution"}
-{"query": "camera conflicts with screen capture", "expected_entities": ["fact:camera-conflict", "fact:coremediaio"], "category": "error-resolution"}
-{"query": "audio gain not applied in pipeline", "expected_entities": ["fact:audio-gain"], "category": "bug-fix"}
-{"query": "Flutter ProviderNotFoundException in secondary window", "expected_entities": ["fact:flutter-provider", "fact:multi-window"], "category": "error-resolution"}
-{"query": "user prefers concise Telegram messages", "expected_entities": ["fact:telegram-preference"], "category": "user-preference"}
-{"query": "PyObjC performRequests_error_ returns bool not tuple", "expected_entities": ["fact:pyobjc-api"], "category": "bug-fix"}
-{"query": "ScreenCaptureKit zero-copy IOSurface", "expected_entities": ["fact:sck-capture", "fact:iosurface"], "category": "tool-knowledge"}
-{"query": "OpenClaw gateway workspace not initialized", "expected_entities": ["fact:workspace-init"], "category": "error-resolution"}
-{"query": "react-native metro bundler cache invalidation", "expected_entities": ["fact:react-native-metro"], "category": "tool-knowledge"}
-{"query": "sinain agent session key format", "expected_entities": ["fact:session-key"], "category": "tool-knowledge"}
-{"query": "what was the OCR backend last month", "expected_entities": ["fact:ocr-backend"], "category": "temporal"}
-{"query": "when did we switch from CGDisplayCreateImage to ScreenCaptureKit", "expected_entities": ["fact:sck-capture", "fact:cgdisplay-deprecation"], "category": "temporal"}