MaaS API

• Google Generative AI API
  • https://ai.google.dev/api/rest
  • Google AI Studio, Gemini API
  • BaseURL https://generativelanguage.googleapis.com/v1beta
  • https://generativelanguage.googleapis.com/v1beta/openai - OpenAI API compatible
  • 接口
    • /models/{model_id}:streamGenerateContent
    • /models/{model_id}:generateContent?key={key}
• Google Vertex AI
  • GCP, 企业级、全托管的机器学习 (ML) 平台
  • Model Garden + MLOps
• Anthropic
  • https://platform.claude.com/docs/en/api/overview
  • SDK
    • https://github.com/anthropics/anthropic-sdk-typescript
    • https://github.com/anthropics/anthropic-sdk-typescript/blob/main/src/resources/beta/messages/messages.ts
• OpenAI
  • https://platform.openai.com/docs/api-reference/chat/create
• XAI
  • https://docs.x.ai/docs/api-reference
• https://www.postman.com/postman/anthropic-apis/documentation/dhus72s/claude-api
• vLLM
  • https://docs.vllm.ai/en/v0.10.2/api/vllm/entrypoints/openai/serving_completion.html
• Provider
  • https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-partner-models
• 错误码
  • https://support.huaweicloud.com/usermanual-maas-modelarts/maas-modelarts-0080.html
• 参考
  • https://www.openresponses.org/

openai	anthropic	google
parallel_tool_calls	disable_parallel_tool_use

• “长尾分布”
• “突发性”
• "Fat Tail" (肥尾)
• 3+Sigma + 15-30min 窗口检查异动

Gemini API

• https://ai.google.dev/api/rest

Multiple tools are supported only when they are all search tools

• 内置 tool 和 functionDeclaration 工具不能同时使用
• openai 里的 tool 映射为一个 functionDeclaration
• 其他的 tool 是内置 tool，语义上有点区别

streaming

• 第一个 chunk 和最后一个 chunk 不应该包含 content
• 有些供应商在第二个 chunk 返回 role
• stream_options
  • continuous_usage_stat
    • 连续发送 usage
  • include_usage
    • 最后一个 chunk 包含 usage
• 参考
  • https://github.com/BerriAI/litellm/blob/4a8629ce/tests/local_testing/test_streaming.py

last chunk

• vLLM, OpenAI 最后一个 chunk 的 content 为 空

{
  "index": 0,
  "delta": {
    "content": ""
  },
  "logprobs": null,
  "finish_reason": "stop",
  "stop_reason": null
}

• 参考
  • https://github.com/BerriAI/litellm/issues/12417
    • LiteLLM 添加最后一个 chunk 的 content 为 空

ToolChoice

• auto
  • 自动选择工具
• required
  • 必须使用工具
• none
  • 不使用工具

Thinking

• https://ai.google.dev/gemini-api/docs/thinking
  • https://ai.google.dev/gemini-api/docs/thinking#set-budget
  • 不同模型支持逻辑不一样
  • 2.5 Pro 不能关闭 128 to 32768
  • 关闭 thinkingBudget = 0
  • 动态 thinkingBudget = -1
  • thinkingLevel
    • 默认 high
    • low, high
    • Gemini 3.0

{
  "contents": [
    {
      "parts": [
        {
          "text": "Provide a list of 3 famous physicists and their key contributions"
        }
      ]
    }
  ],
  "generationConfig": {
    "thinkingConfig": {
      "thinkingLevel": "low"
    }
  }
}

role

• developer
• system
• user
• assistant
• tool
  • 新版本 openai
  • Anthropic 使用 user role
• function
  • 旧版本 openai

Gemini

Missing thought_signature in function call

Please ensure that the number of function response parts is equal to the number of function call parts of the function call turn.

• 非常多类似的错误
• https://github.com/google-gemini/gemini-cli/issues?q=Please%20ensure%20that%20the%20number%20of%20function%20response%20parts%20is%20equal%20to%20the%20number%20of%20function%20call%20parts%20of%20the%20function%20call%20turn.

FAQ

role developer vs system

• OpenAI o1-2024-12-17 之后推出的
• developer 权重比 system 高
• developer
  • 强调规则
• system
  • 强调角色

Unable to submit request because thinking_budget and thinking_level are not supported together

Gemini 限制

Claude temperature, top_p 不能一起传

• Claude Sonnet 4.5 and Claude Haiku 4.5 only support specification of one of temperature or top_p parameters, but cannot handle both.
• 思考与 temperature、top_p 或 top_k 修改不兼容，也不兼容强制使用工具。
• 启用思考后，您无法预先填写响应。
• 对思考预算进行更改，会导致包含消息的缓存提示前缀失效。但是，当思考的参数发生变化时，缓存系统提示和工具定义将继续起作用。
• 参考
  • https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages-request-response.html
  • https://docs.aws.amazon.com/zh_cn/bedrock/latest/userguide/claude-messages-extended-thinking.html