""" 混合路由节点模块 - 前置路由决策 负责决定走快速路径还是 React 循环 """ import re import json from typing import Optional from dataclasses import dataclass, field from datetime import datetime from langchain_core.runnables.config import RunnableConfig from ..state import MainGraphState from backend.app.logger import info, debug from ...model_services.chat_services import get_small_llm_service from ._utils import dispatch_custom_event from backend.app.core.json_parser import extract_and_parse_json, safe_get, safe_get_float, safe_get_str # ========== 核心数据类型 ========== @dataclass class HybridRouterResult: """混合路由结果""" intent: str = "complex" # chitchat / knowledge / tool / complex confidence: float = 0.0 suggested_tools: list = field(default_factory=list) path: str = "react_loop" # fast_chitchat / fast_rag / fast_tool / react_loop reasoning: str = "" # ========== 规则配置 ========== CHITCHAT_KEYWORDS = { "你好", "您好", "hi", "hello", "hey", "早上好", "晚上好", "下午好", "谢谢", "感谢", "多谢", "thanks", "thank you", "再见", "拜拜", "goodbye", "bye" } SUBGRAPH_KEYWORDS = { "contact": ["通讯录", "联系人", "contact", "email", "邮件", "邮箱"], "dictionary": ["词典", "单词", "翻译", "dictionary", "translate", "生词"], "news_analysis": ["资讯", "新闻", "分析", "news", "report", "热点"] } # ========== 意图分类 Prompt 模板 ========== INTENT_CLASSIFICATION_PROMPT = """你是一个专业的意图分类助手。请分析用户的查询，并输出 JSON 格式的结果。 【格式要求】 你必须严格输出 JSON 格式，不要加任何 Markdown 代码块标记（如 ```json）。 仅输出纯 JSON 字符串，不要有其他解释文字。 【意图类型（4选一）： - chitchat: 闲聊、问候、感谢、道别（不需要工具） - knowledge: 知识查询（需要查询知识库） - tool: 工具操作（需要调用通讯录/词典/新闻等子图） - complex: 复杂任务（多步骤、不确定、或需要推理） 【输出格式】 {{ "intent": "chitchat|knowledge|tool|complex", "confidence": 0.85, "reasoning": "简要说明理由", "suggested_tools": ["contact|dictionary|news_analysis", "other"] }} 【重要提示】 - 如果不能100%确定意图，请选择 "complex"，置信度设低一些。 - confidence 是你对当前分类的信心（0.0-1.0）。 - suggested_tools 仅在 intent=tool 时提供，否则设为空数组。 【用户查询】 {query} 【现在开始】 请根据以上信息，输出你的分类 JSON：""" # ========== 规则分流（<5ms） ========== def _rule_based_redirect(query: str) -> Optional[HybridRouterResult]: """规则分流：处理明显不需要推理的情况""" query_clean = query.strip().lower() # 1. 闲聊 if query_clean in CHITCHAT_KEYWORDS or any(kw in query_clean for kw in CHITCHAT_KEYWORDS): return HybridRouterResult( intent="chitchat", confidence=1.0, path="fast_chitchat", reasoning="规则匹配：闲聊类请求" ) # 2. 子图关键词 for subgraph_name, keywords in SUBGRAPH_KEYWORDS.items(): if any(kw in query_clean for kw in keywords): return HybridRouterResult( intent="tool", confidence=0.9, suggested_tools=[subgraph_name], path="fast_tool", reasoning=f"规则匹配：{subgraph_name} 子图关键词" ) # 3. 短问题 if len(query_clean) < 3 or (query_clean.endswith("?") and len(query_clean) < 5): return HybridRouterResult( intent="complex", confidence=0.3, path="react_loop", reasoning="规则匹配：问题过于简短" ) return None # ========== LLM 分类 ========== async def _classify_with_llm(query: str) -> HybridRouterResult: """使用轻量级 LLM 进行意图分类""" try: llm = get_small_llm_service() prompt = INTENT_CLASSIFICATION_PROMPT.format(query=query) response = await llm.ainvoke(prompt) # 使用新的 JSON 解析器 parse_result = extract_and_parse_json(response.content) if not parse_result.success or not parse_result.data: return _default_result() return _parse_classification_result(parse_result.data) except Exception as e: debug(f"LLM 分类失败: {e}") return _default_result() def _parse_classification_result(data: dict) -> HybridRouterResult: """解析分类结果""" intent = safe_get_str(data, "intent", "complex") confidence = safe_get_float(data, "confidence", 0.3) suggested_tools = safe_get(data, "suggested_tools", []) reasoning = safe_get_str(data, "reasoning", "") # 置信度低于阈值，走 complex if confidence < 0.5: intent = "complex" # intent -> path 映射 path_map = { "chitchat": "fast_chitchat", "knowledge": "fast_rag", "tool": "fast_tool", } return HybridRouterResult( intent=intent, confidence=confidence, suggested_tools=suggested_tools, path=path_map.get(intent, "react_loop"), reasoning=reasoning ) def _default_result() -> HybridRouterResult: """默认结果（LLM 失败时）""" return HybridRouterResult( intent="complex", confidence=0.3, path="react_loop", reasoning="LLM 调用失败，降级到 React 循环" ) # ========== 主路由节点 ========== async def hybrid_router_node(state: MainGraphState, config: Optional[RunnableConfig] = None) -> MainGraphState: """混合路由节点：前置路由，决定走快速路径还是 React 循环""" state.current_phase = "hybrid_router" query = state.user_query or "" info(f"[Hybrid Router] 开始路由: {query[:50]}...") # 1. 规则分流 rule_result = _rule_based_redirect(query) if rule_result: decision = rule_result info(f"[Hybrid Router] 规则命中: {decision.path}") else: # 2. LLM 分类 info("[Hybrid Router] 规则未命中，使用 LLM 分类") decision = await _classify_with_llm(query) # 3. 更新状态（同时更新旧的 debug_info 和新的结构化字段） state.debug_info["hybrid_decision"] = { "intent": decision.intent, "confidence": decision.confidence, "path": decision.path, "reasoning": decision.reasoning, "suggested_tools": decision.suggested_tools } state.debug_info["hybrid_start_time"] = datetime.now().isoformat() # 更新新的结构化字段 state.hybrid_router.decision = decision state.hybrid_router.start_time = datetime.now().isoformat() # 4. 发送事件 await dispatch_custom_event("intent_classified", { "intent": decision.intent, "confidence": decision.confidence, "reasoning": decision.reasoning, "suggested_tools": decision.suggested_tools }, config) await dispatch_custom_event("path_decision", { "path": decision.path, "intent": decision.intent, "reasoning": decision.reasoning }, config) info(f"[Hybrid Router] 路由决策: {decision.path} (intent={decision.intent}, confidence={decision.confidence})") return state # ========== 条件路由函数 ========== def route_from_hybrid_decision(state: MainGraphState) -> str: """从混合路由决策获取下一步节点""" decision = state.debug_info.get("hybrid_decision", {}) return decision.get("path", "react_loop") def check_fast_path_success(state: MainGraphState) -> str: """检查快速路径是否成功""" if state.debug_info.get("fast_path_failed"): info("[Fast Path Check] 快速路径失败，升级到 React 循环") return "escalate" info("[Fast Path Check] 快速路径成功，进入 llm_call") return "llm_call" # ========== 导出 ========== __all__ = [ "hybrid_router_node", "route_from_hybrid_decision", "check_fast_path_success", "HybridRouterResult", ]