aiagent/backend/app/services/knowledge_extractor.py

"""
知识提取器 - 从 Agent 执行日志中用 LLM 提取可复用知识
"""
from __future__ import annotations

import json
import logging
from typing import Any, Dict, List, Optional

from sqlalchemy import desc

from app.core.database import SessionLocal

logger = logging.getLogger(__name__)

KNOWLEDGE_EXTRACTION_PROMPT = """你是一个知识工程专家。请从以下 Agent 执行记录中提取可复用的经验知识。

执行输入: {input_text}
执行输出: {output_text}
是否成功: {success}
工具调用链: {tool_chain}
迭代次数: {iterations_used}
工具调用总次数: {tool_calls_made}

请提取为以下 JSON 格式（只输出 JSON，不要其他文本）：
{{
    "title": "一句话标题（如：处理MySQL死锁的重试策略）",
    "category": "bug_fix / best_practice / workaround / optimization / insight",
    "tags": ["tag1", "tag2"],
    "situation": "什么场景下适用",
    "solution": "具体解决方案或操作步骤",
    "caveats": "注意事项或已知限制",
    "confidence": 0.0-1.0
}}

如果这条执行记录没有值得沉淀的知识，返回：
{{"skip": true, "reason": "原因"}}
"""


class KnowledgeExtractor:
    """从执行日志中提取知识（使用 LLM）"""

    def __init__(self, llm_model: str = "deepseek-v4-flash"):
        self.llm_model = llm_model

    def _sync_llm_call(self, prompt: str) -> str:
        """同步调用 LLM。"""
        from app.agent_runtime.core import _LLMClient
        from app.agent_runtime.schemas import AgentLLMConfig
        import asyncio as aio

        client = _LLMClient(AgentLLMConfig(
            provider="deepseek",
            model=self.llm_model,
            temperature=0.3,
            max_iterations=1,
        ))

        result = aio.run(client.chat(
            messages=[{"role": "user", "content": prompt}],
            tools=None,
            iteration=1,
        ))
        content = result.get("content", "") if isinstance(result, dict) else str(result)
        return content

    def extract_from_execution(self, log: Dict[str, Any]) -> Optional[Dict[str, Any]]:
        """从单条执行日志提取知识（同步）。"""
        input_text = log.get("input_text", "") or ""
        output_text = log.get("output_text", "") or ""
        success = log.get("success", False)
        tool_chain = log.get("tool_chain") or []
        tool_calls_made = log.get("tool_calls_made", 0)
        iterations_used = log.get("iterations_used", 0)

        if len(input_text) < 20 or len(output_text) < 50:
            return None
        if not success and not tool_chain:
            return None

        tool_chain_str = json.dumps(tool_chain[:5], ensure_ascii=False) if tool_chain else "无"

        prompt = KNOWLEDGE_EXTRACTION_PROMPT.format(
            input_text=input_text[:2000],
            output_text=output_text[:2000],
            success="是" if success else "否",
            tool_chain=tool_chain_str,
            iterations_used=iterations_used,
            tool_calls_made=tool_calls_made,
        )

        try:
            llm_output = self._sync_llm_call(prompt)
        except Exception as e:
            logger.warning("知识提取 LLM 调用失败: %s", e)
            return None

        try:
            json_start = llm_output.find("{")
            json_end = llm_output.rfind("}") + 1
            if json_start >= 0 and json_end > json_start:
                data = json.loads(llm_output[json_start:json_end])
            else:
                return None
        except json.JSONDecodeError:
            logger.warning("知识提取 JSON 解析失败: %s", llm_output[:200])
            return None

        if data.get("skip"):
            return None

        confidence = float(data.get("confidence", 0.5))
        if confidence < 0.3:
            return None

        return {
            "title": data.get("title", "未命名知识"),
            "category": data.get("category", "insight"),
            "tags": data.get("tags", []),
            "situation": data.get("situation", ""),
            "solution": data.get("solution", ""),
            "caveats": data.get("caveats", ""),
            "confidence": confidence,
        }

    def run_extraction_pipeline(self, limit: int = 10) -> int:
        """运行提取管道：从未处理的执行日志中提取知识。"""
        from app.models.agent_execution_log import AgentExecutionLog
        from app.models.knowledge_entry import KnowledgeEntry

        db = None
        extracted = 0

        try:
            db = SessionLocal()
            logs = (
                db.query(AgentExecutionLog)
                .filter(
                    AgentExecutionLog.success == True,
                    AgentExecutionLog.knowledge_extracted == False,
                    AgentExecutionLog.output_text.isnot(None),
                )
                .order_by(desc(AgentExecutionLog.created_at))
                .limit(limit)
                .all()
            )

            if not logs:
                logger.info("知识提取管道: 没有待处理的新日志")
                return 0

            logger.info("知识提取管道: 找到 %d 条待处理日志", len(logs))

            for log_entry in logs:
                log_data = {
                    "input_text": log_entry.input_text,
                    "output_text": log_entry.output_text,
                    "success": log_entry.success,
                    "tool_chain": log_entry.tool_chain,
                    "iterations_used": log_entry.iterations_used,
                    "tool_calls_made": log_entry.tool_calls_made,
                }

                knowledge = self.extract_from_execution(log_data)
                if not knowledge:
                    log_entry.knowledge_extracted = True
                    db.commit()
                    continue

                existing = (
                    db.query(KnowledgeEntry)
                    .filter(KnowledgeEntry.title == knowledge["title"])
                    .first()
                )
                if existing:
                    log_entry.knowledge_extracted = True
                    db.commit()
                    continue

                embedding_text = (
                    knowledge["title"] + "\n" +
                    knowledge["situation"] + "\n" +
                    knowledge["solution"] + "\n" +
                    knowledge["caveats"]
                )

                entry = KnowledgeEntry(
                    title=knowledge["title"],
                    category=knowledge["category"],
                    tags=knowledge["tags"],
                    situation=knowledge["situation"],
                    solution=knowledge["solution"],
                    caveats=knowledge["caveats"],
                    source_execution_ids=[str(log_entry.id)],
                    source_agent_name=log_entry.agent_name,
                    source_model=log_entry.model,
                    embedding_text=embedding_text,
                    extracted_by="llm_auto",
                    confidence=knowledge["confidence"],
                )
                db.add(entry)
                log_entry.knowledge_extracted = True
                db.commit()
                extracted += 1
                logger.info(
                    "知识提取: %s -> %s (conf=%.2f)",
                    log_entry.agent_name, knowledge["title"], knowledge["confidence"],
                )

            return extracted

        except Exception as e:
            logger.error("知识提取管道异常: %s", e)
            if db:
                try:
                    db.rollback()
                except Exception:
                    pass
            return 0
        finally:
            if db:
                try:
                    db.close()
                except Exception:
                    pass


knowledge_extractor = KnowledgeExtractor()