// ai/router/service.go - 当前实现
func (s *routerServiceImpl) ClassifyIntent(input string, history []string) (*RoutingDecision, error) {
    // Layer 1: Cache
    // Layer 2: Rule matching
    // Layer 3: History matching
    // Layer 4: LLM fallback (耗时 ~100-200ms, 消耗 tokens)
    return s.llmClassify(input, history)
}

type TinyClassifier struct {
    model     *onnx.Model  // ONNX 格式的小模型
    tokenizer *Tokenizer
    threshold float32      // 置信度阈值 (建议 0.85)
}

func (c *TinyClassifier) Predict(input string) (AgentType, Intent, float32) {
    // 1. 分词
    tokens := c.tokenizer.Encode(input)
    
    // 2. 推理
    output := c.model.Run(tokens)
    
    // 3. 解析结果
    agentProbs := output[:5]   // 5 种 AgentType
    intentProbs := output[5:]  // 6 种 Intent
    
    agentIdx, agentConf := argmax(agentProbs)
    intentIdx, intentConf := argmax(intentProbs)
    
    return AgentType(agentIdx), Intent(intentIdx), min(agentConf, intentConf)
}

func (s *routerServiceImpl) ClassifyIntent(input string, history []string) (*RoutingDecision, error) {
    // Layer 1: Cache (不变)
    
    // Layer 2: Rule matching (不变)
    
    // Layer 3: ML Classifier (新增)
    if s.mlClassifier != nil {
        agent, intent, conf := s.mlClassifier.Predict(input)
        if conf >= s.mlThreshold { // 0.85
            return &RoutingDecision{
                AgentType:  agent,
                Intent:     intent,
                Confidence: conf,
            }, nil
        }
    }
    
    // Layer 4: History matching (不变)
    
    // Layer 5: LLM fallback (不变)
    return s.llmClassify(input, history)
}