什么是信号驱动决策？
信号驱动决策是核心架构，它通过从请求中提取多种信号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()
信号驱动路由使用多种信号：
# 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：多个信号共同投票比任何单一信号做出更准确的决策。
6 种信号类型
1. 关键词信号 (Keyword Signals)
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. 嵌入信号 (Embedding Signals)
内容：使用嵌入的语义相似度 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. 领域信号 (Domain Signals)
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. 事实核查信号 (Fact Check Signals)
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. 用户反馈信号 (User Feedback Signals)
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. 偏好信号 (Preference Signals)
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
信号如何组合
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