一个格式对接多家 AI 后端：Azure APIM 统一模型 API 与 MCP 安全策略

# 1. 先确认你的 APIM 实例已启用 Unified Model API 功能
#    在 Azure Portal > API Management > APIs 中找到 "Unified Model API"
#    或通过 REST API 创建：

# 获取 APIM 实例信息
APIM_NAME="my-apim-instance"
RG="my-resource-group"
API_VERSION="2025-06-01"

# 2. 配置后端凭据——把 Anthropic API Key 存为 APIM named value
az rest --method put \
  --url "/subscriptions/{subscriptionId}/resourceGroups/${RG}/providers/Microsoft.ApiManagement/service/${APIM_NAME}/namedValues/anthropic-api-key?api-version=${API_VERSION}" \
  --body '{
    "properties": {
      "displayName": "anthropic-api-key",
      "value": "sk-ant-xxxxxxxxxxxx",
      "secret": true
    }
  }'

# 3. 客户端调用统一格式——注意请求体是 OpenAI 兼容结构
curl -X POST \
  "https://${APIM_NAME}.azure-api.net/unified/chat/completions" \
  -H "Content-Type: application/json" \
  -H "api-key: {你的APIM订阅Key}" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "messages": [
      {"role": "system", "content": "你是一位资深后端工程师"},
      {"role": "user", "content": "用 Python 写一个 FastAPI 健康检查端点"}
    ],
    "max_tokens": 1024,
    "temperature": 0.2
  }'

<!-- APIM inbound policy 示例：按 model 字段路由到不同后端 -->
<policies>
  <inbound>
    <choose>
      <when condition="@(context.Request.Body.As<JObject>()["model"].ToString().StartsWith("claude"))">
        <set-backend-service backend-id="anthropic-backend" />
        <!-- APIM 自动完成格式转换和鉴权注入 -->
      </when>
      <when condition="@(context.Request.Body.As<JObject>()["model"].ToString().StartsWith("gemini"))">
        <set-backend-service backend-id="vertex-ai-backend" />
      </when>
      <otherwise>
        <set-backend-service backend-id="azure-openai-backend" />
      </otherwise>
    </choose>
  </inbound>
</policies>

<!-- APIM 内容安全策略：覆盖 LLM、MCP tool call、A2A 三类流量 -->
<policies>
  <inbound>
    <azure-ai-content-filter>
      <!-- 检测范围扩展 -->
      <traffic-types>
        <type>llm-completion</type>
        <type>mcp-tool-call</type>
        <type>agent-to-agent</type>
      </traffic-types>
      <!-- 对 MCP 工具调用参数中的指令注入进行拦截 -->
      <categories>
        <category name="prompt-injection" severity="medium" action="block" />
        <category name="hate" severity="low" action="annotate" />
        <category name="self-harm" severity="medium" action="block" />
      </categories>
      <!-- 对 A2A payload 中的敏感数据标记 -->
      <sensitive-data>
        <detection>credit-card-number</detection>
        <detection>ssn</detection>
        <action>annotate</action>
      </sensitive-data>
    </azure-ai-content-filter>
  </inbound>
</policies>

# 查询过去 24 小时各后端的 reasoning token 消耗
az monitor metrics list \
  --resource "/subscriptions/{sub}/resourceGroups/${RG}/providers/Microsoft.ApiManagement/service/${APIM_NAME}" \
  --metric "TokenConsumption" \
  --dimension "BackendId" \
  --filter "TokenType eq 'reasoning'" \
  --interval PT1H \
  --start-time $(date -u -d '-24 hours' +%Y-%m-%dT%H:%M:%SZ) \
  --end-time $(date -u +%Y-%m-%dT%H:%M:%SZ)