推理优化

backend/app/main_graph/main_graph_builder.py

@@ -21,7 +21,7 @@ from .nodes.fast_paths import (
     fast_tool_node,
 )
 from .nodes.llm_call import create_dynamic_llm_call_node
-from .nodes.rag_nodes import rag_retrieve_node, check_rag_confidence
+from .nodes.rag_nodes import rag_retrieve_node
 from .nodes.retrieve_memory import create_retrieve_memory_node
 from .nodes.memory_trigger import memory_trigger_node, set_mem0_client
 from .nodes.summarize import create_summarize_node
@@ -164,7 +164,7 @@ def _add_routing_edges(graph: StateGraph, use_hybrid_router: bool, llm_node) ->
             }
         )
 
-        # 快速路径的完成检查
+        # 快速路径的完成检查（fast_rag 失败直接走 react_reason）
         for fast_node in ["fast_chitchat", "fast_rag", "fast_tool"]:
             graph.add_conditional_edges(
                 fast_node,
@@ -198,17 +198,8 @@ def _add_react_loop_edges(graph: StateGraph, subgraph_nodes: Dict[str, Any]) ->
         }
     )
 
-    # RAG 检索后的置信度判断分支
-    graph.add_conditional_edges(
-        "rag_retrieve",
-        check_rag_confidence,
-        {
-            "high_confidence": "llm_call",      # 高置信度 → 直接生成回答
-            "retry_rag": "rag_retrieve",         # 低置信度 → 再次检索
-            "low_confidence": "web_search",      # 两次RAG后仍低 → 联网搜索
-            "no_rag": "web_search",             # 无结果 → 联网搜索
-        }
-    )
+    # RAG 检索后回到 react_reason，由意图识别决定下一步
+    graph.add_edge("rag_retrieve", "react_reason")
 
     # 循环边（回到 react_reason）
     loop_back_nodes = ["web_search", "handle_error"] + subgraph_names