zqf
b1f93588f1
- 排除 .env / *.bak / 内部运维文档（README_INTERNAL.html） - config.py 默认密码已替换为占位符 CHANGE_ME_* - init_db.sql 移除生产数据库用户 GRANT 段 - README.html 数据库用户名已脱敏 - 保留：源码 + 公网 API 文档 + 建表 SQL（无授权语句）
102 lines
3.7 KiB
Python
102 lines
3.7 KiB
Python
"""Fact Extractor - Extract structured atomic facts from working memory using LLM."""
|
||
import json
|
||
import logging
|
||
from typing import Optional, List
|
||
|
||
from lifecycle.llm_parse import parse_llm_json_array
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
FACT_SYSTEM_PROMPT = """你是JSON生成器。输入工作记忆，直接提取结构化事实。
|
||
|
||
严格规则：
|
||
1. 禁止输出分析、推理、思考、解释
|
||
2. 第一个字符必须是 [，最后一个字符必须是 ]
|
||
3. 每条事实：content（50-200字）、importance（0.0-1.0）、category（分类）
|
||
4. 最多20条，无值得提取的事实返回 []
|
||
|
||
输出格式（只输出JSON，不输出其他任何文字）：
|
||
[{"content":"事实","importance":0.8,"category":"分类"}]"""
|
||
|
||
|
||
class FactExtractor:
|
||
"""Extract atomic facts from working memory using LLM."""
|
||
|
||
def __init__(self, mysql_store, redis_cache, llm_client, embedding_service):
|
||
self.mysql = mysql_store
|
||
self.redis = redis_cache
|
||
self.llm = llm_client
|
||
self.embedder = embedding_service
|
||
|
||
def extract(self, agent_id: str, team_id: str,
|
||
max_memories: int = 20,
|
||
delete_after: bool = False) -> dict:
|
||
"""Extract structured facts from working memory."""
|
||
if not self.llm.available:
|
||
return {"error": "LLM not configured. Set LLM_API_URL, LLM_API_KEY, LLM_MODEL environment variables."}
|
||
|
||
items = self.redis.get_working_memories(agent_id, limit=max_memories)
|
||
if not items:
|
||
return {"facts": [], "count": 0}
|
||
|
||
memory_lines = []
|
||
for i, item in enumerate(items):
|
||
if not isinstance(item, dict):
|
||
logger.warning('Skipping non-dict working memory item: %r', item)
|
||
continue
|
||
content = item.get("content", "")[:300]
|
||
memory_lines.append(f"[{i+1}] {content}")
|
||
user_prompt = "以下是需要提取事实的工作记忆：\n\n" + "\n".join(memory_lines)
|
||
|
||
try:
|
||
facts = parse_llm_json_array(
|
||
self.llm, FACT_SYSTEM_PROMPT, user_prompt,
|
||
temperature=0.2, context="extract-facts",
|
||
)
|
||
except ValueError as e:
|
||
return {"error": str(e)}
|
||
|
||
stored = []
|
||
for fact in facts:
|
||
if not isinstance(fact, dict):
|
||
logger.warning('Skipping non-dict fact from LLM: %r', fact)
|
||
continue
|
||
content = fact.get("content", "").strip()
|
||
if not content or len(content) < 10:
|
||
continue
|
||
|
||
importance = fact.get("importance", 0.6)
|
||
if not isinstance(importance, (int, float)) or importance < 0 or importance > 1:
|
||
importance = 0.6
|
||
|
||
category = fact.get("category", "")
|
||
|
||
embedding = self.embedder.embed(content)
|
||
from storage.vector_search import embedding_to_bytes
|
||
emb_bytes = embedding_to_bytes(embedding)
|
||
|
||
memory_id = self.mysql.add_personal_memory(
|
||
agent_id=agent_id,
|
||
team_id=team_id,
|
||
content=content,
|
||
embedding=emb_bytes,
|
||
importance=float(importance),
|
||
metadata={
|
||
"source": "fact_extraction",
|
||
"category": category,
|
||
"extracted_from": "working_memory",
|
||
},
|
||
)
|
||
stored.append({
|
||
"id": memory_id,
|
||
"content": content,
|
||
"importance": importance,
|
||
"category": category,
|
||
})
|
||
|
||
if delete_after and stored:
|
||
self.redis.clear_working_memory(agent_id)
|
||
logger.info(f"Deleted working memories for {agent_id} after extracting {len(stored)} facts")
|
||
|
||
return {"facts": stored, "count": len(stored)}
|