Azure APIM 统一模型 API：用一个接口打通 Anthropic、Vertex AI 等多家后端

# 创建 Anthropic 后端
az apim backend create \
  --resource-group myAIGroup \
  --service-name myAPIM \
  --backend-id anthropic-claude \
  --url "https://api.anthropic.com" \
  --description "Anthropic Claude backend"

# 创建 Vertex AI 后端（备用）
az apim backend create \
  --resource-group myAIGroup \
  --service-name myAPIM \
  --backend-id vertex-gemini \
  --url "https://us-central1-aiplatform.googleapis.com" \
  --description "Google Vertex AI Gemini backend"

<policies>
  <inbound>
    <!-- 验证客户端订阅密钥 -->
    <authentication-subscription-key header-name="Ocp-Apim-Subscription-Key" />

    <!-- 统一请求体转 Anthropic 格式 -->
    <set-body template="liquid">
      {
        "model": "{{body.model | default: 'claude-sonnet-4-20250514'}}",
        "max_tokens": {{body.max_tokens | default: 4096}},
        "messages": [
          {% for msg in body.messages %}
          {
            "role": "{{msg.role}}",
            "content": "{{msg.content}}"
          }{% unless forloop.last %},{% endunless %}
          {% endfor %}
        ]
      }
    </set-body>

    <!-- 注入 Anthropic API Key（存储在 Named Values 中） -->
    <set-header name="x-api-key" exists-action="override">
      <value>{{anthropic-api-key}}</value>
    </set-header>
    <set-header name="anthropic-version" exists-action="override">
      <value>2023-06-01</value>
    </set-header>

    <!-- 内容安全检查：覆盖 LLM 输出 + MCP 工具返回 + A2A payload -->
    <azure-ai-content-safety
      severity-threshold="4"
      categories="Hate Violence SelfHarm Sexual"
      block-on-match="true"
      scan-mcp-tool-response="true"
      scan-a2a-payload="true" />

    <!-- 路由到后端，可按权重切换 -->
    <set-backend-service backend-id="anthropic-claude" />
  </inbound>

  <backend>
    <!-- 限流：每订阅每分钟 60 次 -->
    <limit-call-rate calls="60" renewal-period="60" />
  </backend>

  <outbound>
    <!-- 统一响应格式：把 Anthropic 响应映射回通用结构 -->
    <set-body template="liquid">
      {
        "id": "{{body.id}}",
        "model": "{{body.model}}",
        "choices": [
          {
            "index": 0,
            "message": {
              "role": "{{body.content[0].type}}",
              "content": "{{body.content[0].text}}"
            },
            "finish_reason": "{{body.stop_reason}}"
          }
        ],
        "usage": {
          "prompt_tokens": {{body.usage.input_tokens}},
          "completion_tokens": {{body.usage.output_tokens}},
          "reasoning_tokens": {{body.usage.output_tokens_details.reasoning_tokens | default: 0}},
          "cached_tokens": {{body.usage.input_tokens_details.cache_creation_input_tokens | default: 0}},
          "audio_tokens": 0
        }
      }
    </set-body>

    <!-- 发射 token 指标到 Application Insights -->
    <emit-metric name="LLM-Token-Usage" value="{{body.usage.output_tokens}}" namespace="APIM-LLM">
      <dimension name="Model" value="{{body.model}}" />
      <dimension name="TokenType" value="completion" />
      <dimension name="ReasoningTokens" value="{{body.usage.output_tokens_details.reasoning_tokens | default: 0}}" />
      <dimension name="CachedTokens" value="{{body.usage.input_tokens_details.cache_creation_input_tokens | default: 0}}" />
      <dimension name="AudioTokens" value="0" />
    </emit-metric>
  </outbound>

  <on-error>
    <!-- 错误时也发射指标，便于监控失败率 -->
    <emit-metric name="LLM-Call-Error" value="1" namespace="APIM-LLM">
      <dimension name="ErrorSource" value="{{context.LastError.Source}}" />
      <dimension name="ErrorCode" value="{{context.LastError.Reason}}" />
    </emit-metric>
  </on-error>
</policies>

import httpx

APIM_ENDPOINT = "https://myAPIM.azure-api.net/llm/unified/chat"
APIM_KEY = "你的APIM订阅密钥"

resp = httpx.post(
    APIM_ENDPOINT,
    headers={"Ocp-Apim-Subscription-Key": APIM_KEY},
    json={
        "model": "claude-sonnet-4-20250514",  # 统一模型名，APIM 内部映射到实际后端
        "max_tokens": 1024,
        "messages": [
            {"role": "user", "content": "用三句话解释什么是 MCP 协议"}
        ]
    },
    timeout=30
)

data = resp.json()
print(data["choices"][0]["message"]["content"])
print(f"Token 消耗: prompt={data['usage']['prompt_tokens']}, "
      f"completion={data['usage']['completion_tokens']}, "
      f"reasoning={data['usage']['reasoning_tokens']}")