# agent/citu_agent.py from typing import Dict, Any, Literal from langgraph.graph import StateGraph, END from langchain.agents import AgentExecutor, create_openai_tools_agent from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder from langchain_core.messages import SystemMessage, HumanMessage from agent.state import AgentState from agent.classifier import QuestionClassifier from agent.tools import TOOLS, generate_sql, execute_sql, generate_summary, general_chat from agent.utils import get_compatible_llm class CituLangGraphAgent: """Citu LangGraph智能助手主类 - 使用@tool装饰器 + Agent工具调用""" def __init__(self): # 加载配置 try: from agent.config import get_current_config, get_nested_config self.config = get_current_config() print("[CITU_AGENT] 加载Agent配置完成") except ImportError: self.config = {} print("[CITU_AGENT] 配置文件不可用，使用默认配置") self.classifier = QuestionClassifier() self.tools = TOOLS self.llm = get_compatible_llm() # 注意：现在使用直接工具调用模式，不再需要预创建Agent执行器 print("[CITU_AGENT] 使用直接工具调用模式") self.workflow = self._create_workflow() print("[CITU_AGENT] LangGraph Agent with Direct Tools初始化完成") def _create_workflow(self) -> StateGraph: """创建LangGraph工作流""" workflow = StateGraph(AgentState) # 添加节点 workflow.add_node("classify_question", self._classify_question_node) workflow.add_node("agent_chat", self._agent_chat_node) workflow.add_node("agent_database", self._agent_database_node) workflow.add_node("format_response", self._format_response_node) # 设置入口点 workflow.set_entry_point("classify_question") # 添加条件边：分类后的路由 # 完全信任QuestionClassifier的决策，不再进行二次判断 workflow.add_conditional_edges( "classify_question", self._route_after_classification, { "DATABASE": "agent_database", "CHAT": "agent_chat" # CHAT分支处理所有非DATABASE的情况（包括UNCERTAIN） } ) # 添加边 workflow.add_edge("agent_chat", "format_response") workflow.add_edge("agent_database", "format_response") workflow.add_edge("format_response", END) return workflow.compile() def _classify_question_node(self, state: AgentState) -> AgentState: """问题分类节点""" try: print(f"[CLASSIFY_NODE] 开始分类问题: {state['question']}") classification_result = self.classifier.classify(state["question"]) # 更新状态 state["question_type"] = classification_result.question_type state["classification_confidence"] = classification_result.confidence state["classification_reason"] = classification_result.reason state["classification_method"] = classification_result.method state["current_step"] = "classified" state["execution_path"].append("classify") print(f"[CLASSIFY_NODE] 分类结果: {classification_result.question_type}, 置信度: {classification_result.confidence}") return state except Exception as e: print(f"[ERROR] 问题分类异常: {str(e)}") state["error"] = f"问题分类失败: {str(e)}" state["error_code"] = 500 state["execution_path"].append("classify_error") return state def _agent_database_node(self, state: AgentState) -> AgentState: """数据库Agent节点 - 直接工具调用模式""" try: print(f"[DATABASE_AGENT] 开始处理数据库查询: {state['question']}") question = state["question"] # 步骤1：生成SQL print(f"[DATABASE_AGENT] 步骤1：生成SQL") sql_result = generate_sql.invoke({"question": question, "allow_llm_to_see_data": True}) if not sql_result.get("success"): print(f"[DATABASE_AGENT] SQL生成失败: {sql_result.get('error')}") state["error"] = sql_result.get("error", "SQL生成失败") state["error_code"] = 500 state["current_step"] = "database_error" state["execution_path"].append("agent_database_error") return state sql = sql_result.get("sql") state["sql"] = sql print(f"[DATABASE_AGENT] SQL生成成功: {sql}") # 步骤1.5：检查是否为解释性响应而非SQL error_type = sql_result.get("error_type") if error_type == "llm_explanation": # LLM返回了解释性文本，直接作为最终答案 explanation = sql_result.get("error", "") state["summary"] = explanation + " 请尝试提问其它问题。" state["current_step"] = "database_completed" state["execution_path"].append("agent_database") print(f"[DATABASE_AGENT] 返回LLM解释性答案: {explanation}") return state # 额外验证：检查SQL格式（防止工具误判） from agent.utils import _is_valid_sql_format if not _is_valid_sql_format(sql): # 内容看起来不是SQL，当作解释性响应处理 state["summary"] = sql + " 请尝试提问其它问题。" state["current_step"] = "database_completed" state["execution_path"].append("agent_database") print(f"[DATABASE_AGENT] 内容不是有效SQL，当作解释返回: {sql}") return state # 步骤2：执行SQL print(f"[DATABASE_AGENT] 步骤2：执行SQL") execute_result = execute_sql.invoke({"sql": sql, "max_rows": 200}) if not execute_result.get("success"): print(f"[DATABASE_AGENT] SQL执行失败: {execute_result.get('error')}") state["error"] = execute_result.get("error", "SQL执行失败") state["error_code"] = 500 state["current_step"] = "database_error" state["execution_path"].append("agent_database_error") return state data_result = execute_result.get("data_result") state["data_result"] = data_result print(f"[DATABASE_AGENT] SQL执行成功，返回 {data_result.get('row_count', 0)} 行数据") # 步骤3：生成摘要 print(f"[DATABASE_AGENT] 步骤3：生成摘要") summary_result = generate_summary.invoke({ "question": question, "data_result": data_result, "sql": sql }) if not summary_result.get("success"): print(f"[DATABASE_AGENT] 摘要生成失败: {summary_result.get('message')}") # 摘要生成失败不是致命错误，使用默认摘要 state["summary"] = f"查询执行完成，共返回 {data_result.get('row_count', 0)} 条记录。" else: state["summary"] = summary_result.get("summary") print(f"[DATABASE_AGENT] 摘要生成成功") state["current_step"] = "database_completed" state["execution_path"].append("agent_database") print(f"[DATABASE_AGENT] 数据库查询完成") return state except Exception as e: print(f"[ERROR] 数据库Agent异常: {str(e)}") import traceback print(f"[ERROR] 详细错误信息: {traceback.format_exc()}") state["error"] = f"数据库查询失败: {str(e)}" state["error_code"] = 500 state["current_step"] = "database_error" state["execution_path"].append("agent_database_error") return state def _agent_chat_node(self, state: AgentState) -> AgentState: """聊天Agent节点 - 直接工具调用模式""" try: print(f"[CHAT_AGENT] 开始处理聊天: {state['question']}") question = state["question"] # 构建上下文 enable_context_injection = self.config.get("chat_agent", {}).get("enable_context_injection", True) context = None if enable_context_injection and state.get("classification_reason"): context = f"分类原因: {state['classification_reason']}" # 直接调用general_chat工具 print(f"[CHAT_AGENT] 调用general_chat工具") chat_result = general_chat.invoke({ "question": question, "context": context }) if chat_result.get("success"): state["chat_response"] = chat_result.get("response", "") print(f"[CHAT_AGENT] 聊天处理成功") else: # 处理失败，使用备用响应 state["chat_response"] = chat_result.get("response", "抱歉，我暂时无法处理您的问题。请稍后再试。") print(f"[CHAT_AGENT] 聊天处理失败，使用备用响应: {chat_result.get('error')}") state["current_step"] = "chat_completed" state["execution_path"].append("agent_chat") print(f"[CHAT_AGENT] 聊天处理完成") return state except Exception as e: print(f"[ERROR] 聊天Agent异常: {str(e)}") import traceback print(f"[ERROR] 详细错误信息: {traceback.format_exc()}") state["chat_response"] = "抱歉，我暂时无法处理您的问题。请稍后再试，或者尝试询问数据相关的问题。" state["current_step"] = "chat_error" state["execution_path"].append("agent_chat_error") return state def _format_response_node(self, state: AgentState) -> AgentState: """格式化最终响应节点""" try: print(f"[FORMAT_NODE] 开始格式化响应，问题类型: {state['question_type']}") state["current_step"] = "completed" state["execution_path"].append("format_response") # 根据问题类型和执行状态格式化响应 if state.get("error"): # 有错误的情况 state["final_response"] = { "success": False, "error": state["error"], "error_code": state.get("error_code", 500), "question_type": state["question_type"], "execution_path": state["execution_path"], "classification_info": { "confidence": state.get("classification_confidence", 0), "reason": state.get("classification_reason", ""), "method": state.get("classification_method", "") } } elif state["question_type"] == "DATABASE": # 数据库查询类型 if state.get("summary"): # 有摘要的情况（包括解释性响应和完整查询结果） state["final_response"] = { "success": True, "response": state["summary"], "type": "DATABASE", "sql": state.get("sql"), "data_result": state.get("data_result"), # 可能为None（解释性响应） "summary": state["summary"], "execution_path": state["execution_path"], "classification_info": { "confidence": state["classification_confidence"], "reason": state["classification_reason"], "method": state["classification_method"] } } else: # 数据库查询失败，没有任何结果 state["final_response"] = { "success": False, "error": state.get("error", "数据库查询未完成"), "type": "DATABASE", "sql": state.get("sql"), "execution_path": state["execution_path"] } else: # 聊天类型 state["final_response"] = { "success": True, "response": state.get("chat_response", ""), "type": "CHAT", "execution_path": state["execution_path"], "classification_info": { "confidence": state["classification_confidence"], "reason": state["classification_reason"], "method": state["classification_method"] } } print(f"[FORMAT_NODE] 响应格式化完成") return state except Exception as e: print(f"[ERROR] 响应格式化异常: {str(e)}") state["final_response"] = { "success": False, "error": f"响应格式化异常: {str(e)}", "error_code": 500, "execution_path": state["execution_path"] } return state def _route_after_classification(self, state: AgentState) -> Literal["DATABASE", "CHAT"]: """ 分类后的路由决策 完全信任QuestionClassifier的决策： - DATABASE类型 → 数据库Agent - CHAT和UNCERTAIN类型 → 聊天Agent 这样避免了双重决策的冲突，所有分类逻辑都集中在QuestionClassifier中 """ question_type = state["question_type"] confidence = state["classification_confidence"] print(f"[ROUTE] 分类路由: {question_type}, 置信度: {confidence} (完全信任分类器决策)") if question_type == "DATABASE": return "DATABASE" else: # 将 "CHAT" 和 "UNCERTAIN" 类型都路由到聊天流程 # 聊天Agent可以处理不确定的情况，并在必要时引导用户提供更多信息 return "CHAT" def process_question(self, question: str, session_id: str = None) -> Dict[str, Any]: """ 统一的问题处理入口 Args: question: 用户问题 session_id: 会话ID Returns: Dict包含完整的处理结果 """ try: print(f"[CITU_AGENT] 开始处理问题: {question}") # 初始化状态 initial_state = self._create_initial_state(question, session_id) # 执行工作流 final_state = self.workflow.invoke( initial_state, config={ "configurable": {"session_id": session_id} } if session_id else None ) # 提取最终结果 result = final_state["final_response"] print(f"[CITU_AGENT] 问题处理完成: {result.get('success', False)}") return result except Exception as e: print(f"[ERROR] Agent执行异常: {str(e)}") return { "success": False, "error": f"Agent系统异常: {str(e)}", "error_code": 500, "execution_path": ["error"] } def _create_initial_state(self, question: str, session_id: str = None) -> AgentState: """创建初始状态""" return AgentState( # 输入信息 question=question, session_id=session_id, # 分类结果 question_type="", classification_confidence=0.0, classification_reason="", classification_method="", # 数据库查询流程状态 sql=None, sql_generation_attempts=0, data_result=None, summary=None, # 聊天响应 chat_response=None, # 最终输出 final_response={}, # 错误处理 error=None, error_code=None, # 流程控制 current_step="start", execution_path=[], retry_count=0, max_retries=2, # 调试信息 debug_info={} ) def health_check(self) -> Dict[str, Any]: """健康检查""" try: # 从配置获取健康检查参数 from agent.config import get_nested_config test_question = get_nested_config(self.config, "health_check.test_question", "你好") enable_full_test = get_nested_config(self.config, "health_check.enable_full_test", True) if enable_full_test: # 完整流程测试 test_result = self.process_question(test_question, "health_check") return { "status": "healthy" if test_result.get("success") else "degraded", "test_result": test_result.get("success", False), "workflow_compiled": self.workflow is not None, "tools_count": len(self.tools), "agent_reuse_enabled": False, "message": "Agent健康检查完成" } else: # 简单检查 return { "status": "healthy", "test_result": True, "workflow_compiled": self.workflow is not None, "tools_count": len(self.tools), "agent_reuse_enabled": False, "message": "Agent简单健康检查完成" } except Exception as e: return { "status": "unhealthy", "error": str(e), "workflow_compiled": self.workflow is not None, "tools_count": len(self.tools) if hasattr(self, 'tools') else 0, "agent_reuse_enabled": False, "message": "Agent健康检查失败" }