assistant/agent/agents/router.py

"""
Router Agent - Intent recognition and routing
"""
import json
from typing import Any, Optional

from core.state import AgentState, Intent, ConversationState, set_intent, add_entity
from core.llm import get_llm_client, Message
from utils.logger import get_logger

logger = get_logger(__name__)


# Intent classification prompt
CLASSIFICATION_PROMPT = """你是一个 B2B 购物网站的智能助手路由器。
你的任务是分析用户消息，识别用户意图并提取关键实体。

## 可用意图分类

1. **customer_service** - 通用咨询
   - FAQ 问答
   - 产品使用问题
   - 公司信息查询
   - 政策咨询（退换货政策、隐私政策等）
   
2. **order** - 订单相关
   - 订单查询（"我的订单在哪"、"查一下订单"）
   - 物流跟踪（"快递到哪了"、"什么时候到货"）
   - 订单修改（"改一下收货地址"、"修改订单数量"）
   - 订单取消（"取消订单"、"不想要了"）
   - 发票查询（"开发票"、"要发票"）
   
3. **aftersale** - 售后服务
   - 退货申请（"退货"、"不满意想退"）
   - 换货申请（"换货"、"换一个"）
   - 投诉（"投诉"、"服务态度差"）
   - 工单/问题反馈
   
4. **product** - 商品相关
   - 商品搜索（"有没有xx"、"找一下xx"）
   - 商品推荐（"推荐"、"有什么好的"）
   - 询价（"多少钱"、"批发价"、"大量购买价格"）
   - 库存查询（"有货吗"、"还有多少"）
   
5. **human_handoff** - 需要转人工
   - 用户明确要求转人工
   - 复杂问题 AI 无法处理
   - 敏感问题需要人工处理

## 实体提取

请从消息中提取以下实体（如果存在）：
- order_id: 订单号（如 ORD123456）
- product_id: 商品ID
- product_name: 商品名称
- quantity: 数量
- date_reference: 时间引用（今天、昨天、上周、具体日期等）
- tracking_number: 物流单号
- phone: 电话号码
- address: 地址信息

## 输出格式

请以 JSON 格式返回，包含以下字段：
```json
{
  "intent": "意图分类",
  "confidence": 0.95,
  "sub_intent": "子意图（可选）",
  "entities": {
    "entity_type": "entity_value"
  },
  "reasoning": "简短的推理说明"
}
```

## 注意事项
- 如果意图不明确，置信度应该较低
- 如果无法确定意图，返回 "unknown"
- 实体提取要准确，没有的字段不要填写
"""


async def classify_intent(state: AgentState) -> AgentState:
    """Classify user intent and extract entities
    
    This is the first node in the workflow that analyzes the user's message
    and determines which agent should handle it.
    
    Args:
        state: Current agent state
        
    Returns:
        Updated state with intent and entities
    """
    logger.info(
        "Classifying intent",
        conversation_id=state["conversation_id"],
        message=state["current_message"][:100]
    )
    
    state["state"] = ConversationState.CLASSIFYING.value
    state["step_count"] += 1
    
    # Build context from conversation history
    context_summary = ""
    if state["context"]:
        context_parts = []
        if state["context"].get("order_id"):
            context_parts.append(f"当前讨论的订单: {state['context']['order_id']}")
        if state["context"].get("product_id"):
            context_parts.append(f"当前讨论的商品: {state['context']['product_id']}")
        if context_parts:
            context_summary = "\n".join(context_parts)
    
    # Build messages for LLM
    messages = [
        Message(role="system", content=CLASSIFICATION_PROMPT),
    ]
    
    # Add recent conversation history for context
    for msg in state["messages"][-6:]:  # Last 3 turns
        messages.append(Message(role=msg["role"], content=msg["content"]))
    
    # Add current message with context
    user_content = f"用户消息: {state['current_message']}"
    if context_summary:
        user_content += f"\n\n当前上下文:\n{context_summary}"
    
    messages.append(Message(role="user", content=user_content))
    
    try:
        llm = get_llm_client()
        response = await llm.chat(messages, temperature=0.3)
        
        # Parse JSON response
        content = response.content.strip()
        # Handle markdown code blocks
        if content.startswith("```"):
            content = content.split("```")[1]
            if content.startswith("json"):
                content = content[4:]
        
        result = json.loads(content)
        
        # Extract intent
        intent_str = result.get("intent", "unknown")
        try:
            intent = Intent(intent_str)
        except ValueError:
            intent = Intent.UNKNOWN
        
        confidence = float(result.get("confidence", 0.5))
        sub_intent = result.get("sub_intent")
        
        # Set intent in state
        state = set_intent(state, intent, confidence, sub_intent)
        
        # Extract entities
        entities = result.get("entities", {})
        for entity_type, entity_value in entities.items():
            if entity_value:
                state = add_entity(state, entity_type, entity_value)
        
        logger.info(
            "Intent classified",
            intent=intent.value,
            confidence=confidence,
            entities=list(entities.keys())
        )
        
        # Check if human handoff is needed
        if intent == Intent.HUMAN_HANDOFF or confidence < 0.5:
            state["requires_human"] = True
            state["handoff_reason"] = result.get("reasoning", "Intent unclear")
        
        return state
        
    except json.JSONDecodeError as e:
        logger.error("Failed to parse intent response", error=str(e))
        state["intent"] = Intent.UNKNOWN.value
        state["intent_confidence"] = 0.0
        return state
        
    except Exception as e:
        logger.error("Intent classification failed", error=str(e))
        state["error"] = str(e)
        state["intent"] = Intent.UNKNOWN.value
        return state


def route_by_intent(state: AgentState) -> str:
    """Route to appropriate agent based on intent
    
    This is the routing function used by conditional edges in the graph.
    
    Args:
        state: Current agent state
        
    Returns:
        Name of the next node to execute
    """
    intent = state.get("intent")
    requires_human = state.get("requires_human", False)
    
    # Human handoff takes priority
    if requires_human:
        return "human_handoff"
    
    # Route based on intent
    routing_map = {
        Intent.CUSTOMER_SERVICE.value: "customer_service_agent",
        Intent.ORDER.value: "order_agent",
        Intent.AFTERSALE.value: "aftersale_agent",
        Intent.PRODUCT.value: "product_agent",
        Intent.HUMAN_HANDOFF.value: "human_handoff",
        Intent.UNKNOWN.value: "customer_service_agent"  # Default to customer service
    }
    
    return routing_map.get(intent, "customer_service_agent")