六大Agent模型路由网关横评：成本降85%的秘密架构

# OpenRouter多模型路由示例
import openai

client = openai.OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="sk-or-...",
)

# 简单任务用廉价模型
response = client.chat.completions.create(
    model="google/gemini-3.1-flash",  # $0.10/M tokens
    messages=[{"role": "user", "content": "格式化这段文本"}],
)

# 复杂推理用旗舰模型
response = client.chat.completions.create(
    model="anthropic/claude-opus-4-6",  # $15/M output tokens
    messages=[{"role": "user", "content": "分析这个并购方案的风险"}],
)

# LiteLLM路由配置示例
from litellm import Router

router = Router(
    model_list=[
        {
            "model_name": "agent-cheap",
            "litellm_params": {
                "model": "gemini/gemini-3.1-flash",
                "api_key": "AIza...",
            },
        },
        {
            "model_name": "agent-medium",
            "litellm_params": {
                "model": "gpt-5.4",
                "api_key": "sk-...",
            },
        },
        {
            "model_name": "agent-premium",
            "litellm_params": {
                "model": "claude-opus-4-6",
                "api_key": "sk-ant-...",
            },
        },
    ],
    fallbacks=[{"agent-cheap": ["agent-medium", "agent-premium"]}],
    context_window_fallbacks=[{"agent-cheap": ["agent-medium"]}],
    num_retries=2,
    timeout=30,
    routing_strategy="cost-based",  # 按成本路由
)

# 自动路由到最便宜的可用模型
response = router.completion(
    model="agent-cheap",
    messages=[{"role": "user", "content": "总结这篇文章"}],
)

# Portkey条件路由配置
{
  "strategy": {
    "mode": "conditional",
    "conditions": [
      {
        "query": {
          "metadata.stakes": "high"
        },
        "then": "claude-opus-4-6"
      },
      {
        "query": {
          "metadata.stakes": "medium"
        },
        "then": "gpt-5.4"
      },
      {
        "query": {
          "metadata.stakes": "low"
        },
        "then": "gemini-3.1-flash"
      }
    ],
    "default": "gpt-5.4"
  },
  "fallbacks": [
    {"claude-opus-4-6": ["gpt-5.4"]},
    {"gpt-5.4": ["gemini-3.1-flash"]}
  ]
}

# RouteLLM核心路由逻辑（概念性）
from routellm import Router

router = Router(
    routers=["mf"],  # 矩阵分解路由器
    strong_model="anthropic/claude-opus-4-6",
    weak_model="google/gemini-3.1-flash",
    # 关键：使用真实偏好数据训练
    benchmark="mt-bench",
)

# 路由器自动判断查询复杂度
# 简单查询 → gemini-flash ($0.10/M)
# 复杂查询 → claude-opus ($15/M)
# 结果：85%+ 成本削减，95% 质量保持
response = router.chat.completions.create(
    messages=[{"role": "user", "content": query}],
)

# Semantic Router配置示例
from semantic_router import Route, RouteLayer
from semantic_router.encoders import OpenAIEncoder

encoder = OpenAIEncoder(name="text-embedding-3-small")

# 定义路由规则
coding_route = Route(
    name="coding",
    utterances=[
        "Write a Python function",
        "Debug this code",
        "Refactor this class",
        "Implement a REST API",
    ],
)

analysis_route = Route(
    name="analysis",
    utterances=[
        "Analyze the risk factors",
        "Compare these two approaches",
        "What are the trade-offs",
        "Evaluate the architecture",
    ],
)

simple_route = Route(
    name="simple",
    utterances=[
        "What is the weather",
        "Format this text",
        "Summarize this email",
        "Translate to English",
    ],
)

router = RouteLayer(
    encoder=encoder,
    routes=[coding_route, analysis_route, simple_route],
)

# 路由决策：< 5ms
result = router("Write a function to parse CSV files")
# → Route(name="coding", similarity=0.92)
# → 路由到 GPT-5.4 (最佳代码模型)

# Azure Foundry路由概念性配置
from azure.ai.inference import ModelRouter

router = ModelRouter(
    endpoint="https://your-foundry.services.ai.azure.com",
    # 全局成本预算
    daily_budget_usd=1000,
    # 质量要求
    min_quality_score=0.95,
    # 路由策略
    strategy="cost-aware",  # cost-aware | quality-first | latency-first
)

# 路由器自动做全局优化
# 不是每个查询独立决策，而是考虑今天的预算剩余
response = router.route(
    messages=[{"role": "user", "content": "分析这份合同的风险"}],
    metadata={"stakes": "high", "domain": "legal"},
)
# → 根据当日预算消耗情况动态选择模型
# 预算充裕 → Claude Opus ($15/M)
# 预算紧张 → GPT-5.4 ($8/M) + 严格验证

graph TB
    subgraph "第一层：意图分类 (Intent Classification)"
        A[用户请求] --> B{Semantic Router<br/>嵌入匹配 < 5ms}
        B -->|简单查询<br/>sim > 0.85| C[轻量级模型<br/>Gemini Flash / Llama]
        B -->|复杂查询<br/>sim < 0.85| D{第二层：复杂度评估}
        B -->|模糊查询<br/>无法分类| E[LLM路由器<br/>GPT-4o-mini分类]
    end

    subgraph "第二层：复杂度路由 (Complexity Routing)"
        D --> F[RouteLLM<br/>偏好训练路由器]
        F -->|低复杂度| G[中端模型<br/>GPT-4o / Claude Sonnet]
        F -->|高复杂度| H[旗舰模型<br/>Claude Opus / GPT-5.4]
        E -->|分类结果| F
    end

    subgraph "第三层：Fallback与预算控制 (Fallback & Budget)"
        C -->|失败/超时| G
        G -->|失败/低置信度| H
        H -->|连续3次失败| I[断路器<br/>Portkey Circuit Breaker]
        I --> J[人类接管]
        H -->|成本超预算| K[降级到中端模型<br/>+ 严格验证]
    end

    subgraph "网关层 (Gateway)"
        L[LiteLLM Proxy<br/>统一API格式]
        M[Portkey Gateway<br/>可观测+缓存+Guardrails]
        N[OpenRouter<br/>300+模型市场]
    end

    C --> L
    G --> L
    H --> L
    L --> M
    M --> N

graph LR
    A[100%请求] --> B["Level 1: 本地模型<br/>Ollama Llama 3<br/>$0/M tokens"]
    B -->|60%成功| C[完成 ✅]
    B -->|40%失败| D["Level 2: 廉价云模型<br/>Gemini Flash<br/>$0.10/M tokens"]
    D -->|25%成功| C
    D -->|15%失败| E["Level 3: 平衡旗舰<br/>Claude Sonnet<br/>$3/M output tokens"]
    E -->|10%成功| C
    E -->|5%失败| F["Level 4: 旗舰模型<br/>Claude Opus<br/>$25/M output tokens"]
    F -->|5%成功| C

    style B fill:#4CAF50,color:#fff
    style D fill:#FFC107,color:#000
    style E fill:#FF9800,color:#fff
    style F fill:#F44336,color:#fff

graph TD
    A[开始选型] --> B{日均请求量?}
    B -->|< 1K| C[OpenRouter<br/>无需基础设施]
    B -->|1K-100K| D{是否需要自托管?}
    B -->|> 100K| E{合规要求?}

    D -->|是| F[LiteLLM + RouteLLM]
    D -->|否| G[Portkey托管网关]

    E -->|金融/医疗/政府| F
    E -->|一般商业| H{延迟敏感?}

    H -->|< 50ms| I[Semantic Router<br/>+ LiteLLM Fallback]
    H -->|> 50ms可接受| F

    C --> J{是否需要<br/>可观测性?}
    J -->|是| K[OpenRouter + Helicone]
    J -->|否| C