修改rag，实现混合检索

backend/app/agent/agent_service.py

@@ -70,7 +70,7 @@ class AIAgentService:
             raise RuntimeError("没有可用的模型")
         return self
 
-    async def process_message(self, message: str, thread_id: str, model: str = "zhipu", user_id: str = "default_user") -> dict:
+    async def process_message(self, message: str, thread_id: str, model: str = "local", user_id: str = "default_user") -> dict:
         """处理用户消息，返回包含回复、token统计和耗时的字典"""
         if model not in self.graphs:
             # 回退到第一个可用模型
@@ -175,6 +175,8 @@ class AIAgentService:
         try:
             info(f"📡 开始调用 graph.astream()...")
             chunk_count = 0
+            full_message_content = ""  # 收集完整消息内容
+            
             async for chunk in graph.astream(
                 input_state,
                 config=config,
@@ -184,21 +186,11 @@ class AIAgentService:
             ):
                 chunk_count += 1
                 chunk_type = chunk["type"]
-                info(f"📦 收到第 {chunk_count} 个chunk, type: {chunk_type}")
                 processed_event = {}
 
                 if chunk_type == "messages":
                     message_chunk, metadata = chunk["data"]
                     node_name = metadata.get("langgraph_node", "unknown")
-                    info(f"📨 处理消息chunk, node: {node_name}")
-                    # 详细记录消息内容，看看这些 chunk 到底是什么
-                    if hasattr(message_chunk, "content"):
-                        content_preview = str(message_chunk.content)[:200]
-                        info(f"📄 消息内容预览 ({len(content_preview)} chars): {repr(content_preview)}")
-                    if hasattr(message_chunk, "type"):
-                        info(f"📋 消息类型: {message_chunk.type}")
-                    if hasattr(message_chunk, "tool_calls"):
-                        info(f"🔧 包含工具调用: {message_chunk.tool_calls}")
 
                     # 检测节点变化，发送节点开始事件
                     if node_name != current_node:
@@ -218,8 +210,6 @@ class AIAgentService:
                     reasoning_token = ""
                     if hasattr(message_chunk, 'additional_kwargs'):
                         reasoning_token = message_chunk.additional_kwargs.get("reasoning_content", "")
-                    
-                    info(f"💬 消息token: token_content='{repr(token_content[:50])}', reasoning_token='{repr(reasoning_token[:50])}', node_name='{node_name}'")
 
                     # 处理思考过程
                     if reasoning_token:
@@ -228,7 +218,6 @@ class AIAgentService:
                             "node": node_name,
                             "reasoning_token": reasoning_token
                         }
-                        info(f"✅ 生成 reasoning_token 事件: {processed_event}")
                     # 处理工具调用
                     elif hasattr(message_chunk, 'tool_calls') and message_chunk.tool_calls:
                         for tool_call in message_chunk.tool_calls:
@@ -248,7 +237,7 @@ class AIAgentService:
                                     "args": tool_args,
                                     "id": tool_call_id
                                 }
-                    # 处理普通 token
+                    # 处理普通 token - 只收集，不打印单个 token
                     elif token_content:
                         processed_event = {
                             "type": "llm_token",
@@ -256,18 +245,13 @@ class AIAgentService:
                             "token": token_content,
                             "reasoning_token": reasoning_token
                         }
-                        info(f"✅ 生成 llm_token 事件: {processed_event}")
-                    else:
-                        info(f"⚠️ 没有生成任何事件，token_content='{repr(token_content)}', reasoning_token='{repr(reasoning_token)}'")
+                        if node_name == "llm_call":
+                            full_message_content += token_content
 
                 elif chunk_type == "updates":
-                    info(f"🔄 处理updates chunk")
                     updates_data = chunk["data"]
                     serialized_data = self._serialize_value(updates_data)
                     
-                    # 关键修复：不再从 updates 中读取 latest_reasoning，避免重复
-                    # 因为我们现在直接通过 custom 事件发送推理结果了
-                    
                     # 检查是否有人工审核请求
                     if "review_pending" in serialized_data and serialized_data["review_pending"]:
                         review_id = serialized_data.get("review_id", "")
@@ -302,18 +286,12 @@ class AIAgentService:
                     }
 
                 elif chunk_type == "custom":
-                    info(f"🎯 处理custom chunk, 完整数据: {repr(chunk)}")
                     custom_data = chunk["data"]
-                    info(f"🎯 custom_data 内容: {repr(custom_data)}")
-                    info(f"🎯 custom_data 类型: {type(custom_data)}")
                     
-                    # 关键修复：处理我们从 react_reason_node 发送的自定义推理事件
-                    # LangGraph 的 adispatch_custom_event 发送的事件格式：
-                    # chunk["data"] 是我们传的第二个参数（dict）
+                    # 处理我们从 react_reason_node 发送的自定义推理事件
                     if isinstance(custom_data, dict):
                         # 检查是否是我们的推理事件
                         if "action" in custom_data and "reasoning" in custom_data:
-                            info(f"[Agent Service] 收到自定义推理事件: {custom_data}")
                             yield {
                                 "type": "react_reasoning",
                                 "step": custom_data.get("step", 1),
@@ -339,7 +317,10 @@ class AIAgentService:
                 if processed_event:
                     yield processed_event
 
+            # 完整消息集合完成后，一次性打印
             info(f"✅ graph.astream() 完成，共 {chunk_count} 个chunks")
+            if full_message_content:
+                info(f"📄 完整消息内容: {repr(full_message_content)}")
 
         except Exception as e:
             error(f"❌ 执行 React 图时出错: {e}")