fix: 修复 RAG 无限循环问题和导入错误
Some checks failed
构建并部署 AI Agent 服务 / deploy (push) Has been cancelled

主要修复:
1. 修复 RAG 推理无限循环问题(大小写不匹配 + 缺少已检索结果检查)
2. 修复 intent_classifier.py 的绝对导入错误
3. 删除旧的 start.sh 脚本,添加新的启动脚本
4. 优化路由逻辑和状态管理
This commit is contained in:
2026-05-04 18:59:15 +08:00
parent 9841f47432
commit c9bf21be0e
13 changed files with 503 additions and 164 deletions

View File

@@ -8,10 +8,10 @@ from typing import Dict, Any, Optional, List
from dataclasses import dataclass, field
from datetime import datetime
from app.main_graph.state import MainGraphState
from app.logger import info, debug
from app.model_services.chat_services import get_small_llm_service, get_chat_service
from app.main_graph.nodes.rag_nodes import rag_retrieve_node
from ..state import MainGraphState
from ...logger import info, debug
from ...model_services.chat_services import get_small_llm_service, get_chat_service
from .rag_nodes import rag_retrieve_node
# ========== 核心数据类型 ==========
@@ -367,8 +367,8 @@ async def fast_rag_node(state: MainGraphState, config: Optional[Dict[str, Any]]
debug(f"[Fast RAG] 发送事件失败: {e}")
try:
# 先尝试 RAG 检索
state = rag_retrieve_node(state, config)
# 先尝试 RAG 检索 - 注意rag_retrieve_node 是异步函数,需要 await
state = await rag_retrieve_node(state, config)
# 检查检索结果
rag_docs = getattr(state, "rag_docs", [])

View File

@@ -364,11 +364,15 @@ def route_by_reasoning(state: MainGraphState) -> str:
if "subgraph_completed" in previous_actions or state.final_result:
return "llm_call"
# 关键修复:如果已经执行过 rag_retrieve 并且又执行过推理,直接去 LLM_CALL
# 这样的流程推理1 → RAG → 推理2判断 RAG 结果) → LLM_CALL
rag_count = previous_actions.count("rag_retrieve")
if rag_count >= 1 and len(previous_actions) >= rag_count + 1:
info(f"[route_by_reasoning] 已完成 RAG 检索和结果判断,直接去 llm_call")
# 关键修复:检测 RAG 重复循环 - 如果发现"RETRIEVE_RAG"出现超过1次,直接去 LLM
rag_count = previous_actions.count("RETRIEVE_RAG")
if rag_count >= 2:
info(f"[route_by_reasoning] 检测到 RAG 重复循环({rag_count}次),直接去 llm_call")
return "llm_call"
# 关键修复:如果已经有 rag_docs 或 rag_context说明已经检索过了直接去 LLM
if (state.rag_docs and len(state.rag_docs) > 0) or (state.rag_context and len(state.rag_context) > 0):
info(f"[route_by_reasoning] 检测到已存在 RAG 检索结果,直接去 llm_call")
return "llm_call"
# 关键修复:限制最多 3 次推理,避免无限循环