homegraph 1.0.0 → 1.1.2

package/scripts/qa_eval/eval_metrics.py

@@ -0,0 +1,233 @@
+#!/usr/bin/env python3
+"""
+Stage 2 — RAG Answer Accuracy Evaluation (LLM-as-Judge).
+
+Reads agent-produced JSONL (query + output_answer + reference_answer),
+strips tool-call blocks, scores 0–1 via MyAnswerAccuracy.
+
+Usage:
+  export DASHSCOPE_API_KEY="sk-..."
+  python scripts/qa_eval/eval_metrics.py \\
+    -i scripts/qa_eval/log/result-with-builtin.jsonl \\
+    -o scripts/qa_eval/log/result-with-builtin-scored.jsonl \\
+    -w 2
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import json
+import logging
+import os
+import statistics
+import sys
+import threading
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+
+from tqdm import tqdm
+
+_SCRIPT_DIR = Path(__file__).resolve().parent
+LOG_DIR = _SCRIPT_DIR / "log"
+LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+if str(_SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(_SCRIPT_DIR))
+
+from my_answer_accuracy import MyAnswerAccuracy  # noqa: E402
+from llm_config import PROVIDER_DASHSCOPE, PROVIDER_ZHIPU, provider_help, resolve_llm_config  # noqa: E402
+
+os.environ.setdefault("NO_PROXY", "localhost,127.0.0.1")
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(levelname)s - %(message)s",
+    handlers=[
+        logging.FileHandler(LOG_DIR / "eval_metrics.log"),
+        logging.StreamHandler(),
+    ],
+)
+logger = logging.getLogger(__name__)
+
+file_write_lock = threading.Lock()
+
+DEFAULT_INPUT = LOG_DIR / "result-with-builtin.jsonl"
+DEFAULT_OUTPUT = LOG_DIR / "result-with-builtin-scored.jsonl"
+DEFAULT_MODEL = "qwen3-235b-a22b-instruct-2507"
+DEFAULT_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
+DEFAULT_API_KEY_ENV = "DASHSCOPE_API_KEY"
+
+
+class RAGEvaluator:
+    """Answer accuracy judge with concurrent processing."""
+
+    def __init__(
+        self,
+        llm_model: str,
+        *,
+        api_key: str,
+        base_url: str,
+        extra_body: dict | None = None,
+    ) -> None:
+        self.scorer = MyAnswerAccuracy.create(
+            api_key=api_key,
+            base_url=base_url,
+            model=llm_model,
+            extra_body=extra_body,
+        )
+
+    async def evaluate_answer_accuracy_single(self, sample_data: dict) -> dict:
+        return await self.scorer.evaluate_answer_accuracy_single(sample_data)
+
+
+def process_single_item_sync(evaluator: RAGEvaluator, item: dict) -> dict:
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
+    try:
+        return loop.run_until_complete(evaluator.evaluate_answer_accuracy_single(item))
+    finally:
+        loop.close()
+
+
+def load_jsonl(path: Path) -> list[dict]:
+    items: list[dict] = []
+    with path.open("r", encoding="utf-8") as f:
+        for line_no, line in enumerate(f, 1):
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                items.append(json.loads(line))
+            except json.JSONDecodeError as e:
+                logger.error("第 %s 行 JSON 解析失败: %s", line_no, e)
+    return items
+
+
+def resolve_api_key(env_name: str) -> str | None:
+    key = os.environ.get(env_name, "").strip()
+    if key:
+        return key
+    fallback = os.environ.get("OPENAI_API_KEY", "").strip()
+    return fallback or None
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="RAG Answer Accuracy Evaluation (LLM-as-Judge)")
+    parser.add_argument("--input", "-i", type=str, default=str(DEFAULT_INPUT), help="Agent 产出的 JSONL")
+    parser.add_argument("--output", "-o", type=str, default=str(DEFAULT_OUTPUT), help="打分后的 JSONL")
+    parser.add_argument("--workers", "-w", type=int, default=1, help="并发线程数")
+    parser.add_argument(
+        "--provider",
+        choices=[PROVIDER_DASHSCOPE, PROVIDER_ZHIPU],
+        default=None,
+        help=provider_help(),
+    )
+    parser.add_argument("--model", "-m", type=str, default=None, help="Judge 模型（默认随 provider）")
+    parser.add_argument("--base-url", type=str, default=None, help="OpenAI 兼容 API 端点（默认随 provider）")
+    args = parser.parse_args()
+
+    input_file = Path(args.input)
+    output_file = Path(args.output)
+
+    try:
+        llm = resolve_llm_config(provider=args.provider, model=args.model, base_url=args.base_url)
+    except RuntimeError as e:
+        print(f"错误: {e}", file=sys.stderr)
+        return 1
+
+    print(f"输入文件: {input_file}")
+    print(f"输出文件: {output_file}")
+    print(f"并发线程数: {args.workers}")
+    print(f"LLM provider: {llm.provider}")
+    print(f"评估模型: {llm.model}")
+
+    if not input_file.is_file():
+        print(f"错误: 输入文件 {input_file} 不存在")
+        return 1
+
+    api_key = llm.api_key
+
+    all_items = load_jsonl(input_file)
+    total_items = len(all_items)
+    print(f"读取到 {total_items} 个样本")
+
+    if total_items == 0:
+        print("没有找到有效的样本数据")
+        return 1
+
+    for i, item in enumerate(all_items, 1):
+        if not str(item.get("query", "")).strip():
+            print(f"错误: 第 {i} 条缺少 query")
+            return 1
+        if not str(item.get("reference_answer", "")).strip():
+            print(f"错误: 第 {i} 条缺少 reference_answer")
+            return 1
+        if not str(item.get("output_answer", "")).strip():
+            logger.warning("第 %s 条 (%s) output_answer 为空，Judge 可能得 0 分", i, item.get("id", "?"))
+
+    output_file.parent.mkdir(parents=True, exist_ok=True)
+    output_file.write_text("", encoding="utf-8")
+
+    evaluator = RAGEvaluator(
+        llm.model,
+        api_key=api_key,
+        base_url=llm.base_url,
+        extra_body=llm.extra_body,
+    )
+    print("初始化 Judge 完成")
+
+    processed_items: list[dict] = []
+    results_lock = threading.Lock()
+
+    with ThreadPoolExecutor(max_workers=max(1, args.workers)) as executor:
+        futures = {
+            executor.submit(process_single_item_sync, evaluator, item): item
+            for item in all_items
+        }
+        for future in tqdm(as_completed(futures), total=len(all_items), desc="评估答案准确性"):
+            try:
+                processed_item = future.result()
+                with results_lock:
+                    processed_items.append(processed_item)
+                    if len(processed_items) % 5 == 0:
+                        with file_write_lock:
+                            with output_file.open("a", encoding="utf-8") as writer:
+                                for row in processed_items[-5:]:
+                                    writer.write(json.dumps(row, ensure_ascii=False) + "\n")
+            except Exception as e:
+                logger.error("样本处理任务异常: %s", e)
+
+    remaining = len(processed_items) % 5
+    if remaining:
+        with file_write_lock:
+            with output_file.open("a", encoding="utf-8") as writer:
+                for row in processed_items[-remaining:]:
+                    writer.write(json.dumps(row, ensure_ascii=False) + "\n")
+
+    if processed_items:
+        successful = [x for x in processed_items if x.get("evaluation_status") == "success"]
+        scores = [float(x.get("answer_accuracy_score", 0.0)) for x in processed_items]
+        print("\n评估完成!")
+        print(f"总样本数: {total_items}")
+        print(f"成功评估: {len(successful)}")
+        print(f"失败评估: {total_items - len(successful)}")
+        if successful:
+            ok_scores = [float(x["answer_accuracy_score"]) for x in successful]
+            print(f"平均答案准确性得分: {statistics.mean(ok_scores):.4f}")
+            print(f"最高得分: {max(ok_scores):.4f}")
+            print(f"最低得分: {min(ok_scores):.4f}")
+        print(f"\n结果已写入: {output_file}")
+    else:
+        print("没有成功处理任何样本")
+        return 1
+
+    return 0
+
+
+if __name__ == "__main__":
+    try:
+        raise SystemExit(main())
+    except KeyboardInterrupt:
+        print("\n用户中断了程序执行")
+        raise SystemExit(130)