智能模型路由配置教程

模型路由概述

在多模型环境中，不同类型的请求适合不同的模型。OpenClaw 的智能模型路由可以根据消息内容、用户属性和频道设置自动选择最合适的模型，实现性能与成本的最优平衡。

基础路由配置

{
  "routing": {
    "enabled": true,
    "defaultModel": "fast",
    "rules": [
      {
        "name": "code-tasks",
        "match": {"content": "代码|编程|bug|函数|调试|Python|JavaScript"},
        "model": "code-expert"
      },
      {
        "name": "complex-reasoning",
        "match": {"content": "分析|比较|推理|规划|策略"},
        "model": "smart"
      },
      {
        "name": "translation",
        "match": {"content": "翻译|translate|英译中|中译英"},
        "model": "translator"
      }
    ]
  }
}

匹配条件

按内容匹配

{
  "match": {
    "content": "关键词1|关键词2|关键词3"
  }
}

支持正则表达式：

{
  "match": {
    "content": "\\b(def|class|function|import)\\b"
  }
}

按用户匹配

{
  "match": {
    "user": ["user001", "user002"]
  }
}

按频道匹配

{
  "match": {
    "channel": "discord-dev"
  }
}

按消息长度匹配

{
  "match": {
    "contentLength": {
      "min": 500
    }
  }
}

组合条件

{
  "match": {
    "content": "代码|编程",
    "channel": "discord-dev",
    "user": ["dev-user1", "dev-user2"]
  }
}

多个条件之间是 AND 关系。

完整路由配置示例

{
  "routing": {
    "enabled": true,
    "defaultModel": "fast",
    "rules": [
      {
        "name": "vip-users",
        "priority": 100,
        "match": {"user": ["vip001", "vip002"]},
        "model": "smart"
      },
      {
        "name": "coding-channel",
        "priority": 90,
        "match": {"channel": "discord-dev"},
        "model": "code-expert"
      },
      {
        "name": "long-content",
        "priority": 80,
        "match": {"contentLength": {"min": 500}},
        "model": "smart"
      },
      {
        "name": "code-keywords",
        "priority": 70,
        "match": {"content": "\\b(code|bug|function|class)\\b"},
        "model": "code-expert"
      },
      {
        "name": "default",
        "priority": 0,
        "match": {"content": ".*"},
        "model": "fast"
      }
    ]
  }
}

规则按 priority 从高到低匹配，第一个匹配的规则生效。

AI 分类路由

使用轻量级模型进行意图分类，再路由到合适的模型：

{
  "routing": {
    "classifier": {
      "enabled": true,
      "model": "fast",
      "categories": {
        "coding": {"model": "code-expert"},
        "creative": {"model": "smart"},
        "translation": {"model": "translator"},
        "general": {"model": "fast"}
      }
    }
  }
}

分类器会先用轻量模型判断意图类别，再将请求路由到对应模型。

成本路由

按成本预算进行路由：

{
  "routing": {
    "costAware": {
      "enabled": true,
      "dailyBudget": 20.00,
      "thresholds": [
        {"usage": 0.8, "model": "fast"},
        {"usage": 0.5, "model": "smart"},
        {"usage": 0.0, "model": "premium"}
      ]
    }
  }
}

当日消费低于预算 50% 时使用最好的模型，超过 80% 后自动降级到经济模型。

查看路由统计

openclaw routing stats

Routing Statistics (last 24h):
  Rule                Matches   Model         Avg Latency
  ─────────────────────────────────────────────────────
  vip-users           120       smart         2.1s
  coding-channel      85        code-expert   1.8s
  code-keywords       200       code-expert   1.7s
  default             1,500     fast          0.8s

测试路由规则

openclaw routing test "帮我写一个Python排序函数"

Message: "帮我写一个Python排序函数"
  Matched rule: code-keywords (priority: 70)
  Selected model: code-expert
  Reason: Content matches pattern "\\b(code|bug|function|class)\\b"

总结

智能模型路由让 OpenClaw 能够在多模型环境中做出最优选择。通过基于内容、用户和预算的路由规则，可以在保证服务质量的同时最大化资源利用效率。