什么是 Signal-Driven Decision？

Signal-Driven Decision 是核心架构，它通过从请求中提取多种 signal 并将它们结合起来做出更好的路由决策，从而实现智能路由。

核心理念

传统路由使用单一信号：

# 传统：单一分类模型
if classifier(query) == "math":
    route_to_math_model()

Signal-Driven routing 使用多种 signal：

# 信号驱动：多种信号组合
if (keyword_match AND domain_match) OR high_embedding_similarity:
    route_to_math_model()

为什么这很重要：多个 signal 共同投票比任何单一 signal 做出更准确的决策。

6 种 Signal 类型

1. Keyword Signal

内容：使用 AND/OR 运算符的快速模式匹配 延迟：小于 1ms 用例：确定性路由、合规性、安全性

signals:
  keywords:
    - name: "math_keywords"
      operator: "OR"
      keywords: ["calculate", "equation", "solve", "derivative"]

示例："Calculate the derivative of x^2" → 匹配 "calculate" 和 "derivative"

2. Embedding Signal

内容：使用 embedding 的语义相似度 延迟：10-50ms 用例：意图检测、释义处理

signals:
  embeddings:
    - name: "code_debug"
      threshold: 0.70
      candidates:
        - "My code isn't working, how do I fix it?"
        - "Help me debug this function"

示例："Need help debugging this function" → 0.78 相似度 → 匹配！

3. Domain Signal

内容：MMLU 领域分类（14 个类别） 延迟：50-100ms 用例：学术和专业领域路由

signals:
  domains:
    - name: "mathematics"
      mmlu_categories: ["abstract_algebra", "college_mathematics"]

示例："Prove that the square root of 2 is irrational" → Mathematics (数学) 领域

4. Fact Check Signal

内容：基于机器学习的需要事实验证的查询检测 延迟：50-100ms 用例：医疗保健、金融服务、教育

signals:
  fact_checks:
    - name: "factual_queries"
      threshold: 0.75

示例："What is the capital of France?" → 需要事实核查

5. User Feedback Signal

内容：用户反馈和更正的分类 延迟：50-100ms 用例：客户支持、自适应学习

signals:
  user_feedbacks:
    - name: "negative_feedback"
      feedback_types: ["correction", "dissatisfaction"]

示例："That's wrong, try again" → 检测到负面反馈

6. Preference Signal

内容：基于 LLM 的路由偏好匹配 延迟：200-500ms 用例：复杂意图分析

signals:
  preferences:
    - name: "creative_writing"
      llm_endpoint: "http://localhost:8000/v1"
      model: "gpt-4"
      routes:
        - name: "creative"
          description: "Creative writing, storytelling, poetry"

示例："Write a story about dragons" → 偏好创意路由

Signal 如何组合

AND 运算符 - 必须全部匹配

decisions:
  - name: "advanced_math"
    rules:
      operator: "AND"
      conditions:
        - type: "keyword"
          name: "math_keywords"
        - type: "domain"
          name: "mathematics"

逻辑：仅当关键词 AND (并且) 领域都匹配时，路由到 advanced_math

用例：高置信度路由（减少误报）

OR 运算符 - 任意匹配

decisions:
  - name: "code_help"
    rules:
      operator: "OR"
      conditions:
        - type: "keyword"
          name: "code_keywords"
        - type: "embedding"
          name: "code_debug"

逻辑：如果关键词 OR (或者) 嵌入匹配，路由到 code_help

用例：广泛覆盖（减少漏报）

嵌套逻辑 - 复杂规则

decisions:
  - name: "verified_math"
    rules:
      operator: "AND"
      conditions:
        - type: "domain"
          name: "mathematics"
        - operator: "OR"
          conditions:
            - type: "keyword"
              name: "proof_keywords"
            - type: "fact_check"
              name: "factual_queries"

逻辑：如果 (数学领域) AND (证明关键词 OR 需要事实核查) 则路由

用例：复杂路由场景

真实世界示例

用户查询

"Prove that the square root of 2 is irrational"

信号提取

signals_detected:
  keyword: true          # "prove", "square root", "irrational"
  embedding: 0.89        # 与数学查询的高度相似性
  domain: "mathematics"  # MMLU 分类
  fact_check: true       # 证明需要验证

决策过程

decision: "advanced_math"
reason: "All math signals agree (keyword + embedding + domain + fact_check)" # 所有数学信号一致
confidence: 0.95
selected_model: "qwen-math"

为什么这有效

• 多个信号一致：高置信度
• 启用了事实核查：质量保证
• 专业模型：最适合数学证明

下一步

• 配置指南 - 配置 signal 和 decision
• Keyword Routing 教程 - 学习 keyword signal
• Embedding Routing 教程 - 学习 embedding signal
• Domain Routing 教程 - 学习 domain signal