Skip to main content

Gemini

An llm · client with dialect: gemini sends requests to the Google Gemini API. Requests that arrived in the OpenAI or Anthropic API are translated to the Gemini format and back.

Configuration

gemini_llm_client:
  type: llm
  kind: client
  options:
    dialect: gemini
    server: https://generativelanguage.googleapis.com
    authorization:
      api_key:
        credentials: "{credentials}"
  telemetry:
    metrics:
      - llm.*
  exit: sys:http_client

Server URL

Zilla appends /models/{model}:generateContent, or /models/{model}:streamGenerateContent when the reply is streamed, to the server URL, with {model} taken from the request. When the URL has no path, or only /, the path defaults to /v1beta, so https://generativelanguage.googleapis.com sends requests to https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent. Set a path, such as https://generativelanguage.googleapis.com/v1, to use the stable API instead.

A request that arrived in the Gemini API is sent with the same operation the caller used. A request translated from another API always calls :streamGenerateContent with alt=sse, and the reply is translated as it arrives.

Credentials

The Gemini API takes the API key in the x-goog-api-key header, without a prefix, so set credentials: "{credentials}" instead of the default Bearer {credentials}. With an inline guard, the key is the one the caller sent, forwarded unchanged. See LLM Credentials.

Routing to Gemini

north_llm_proxy:
  type: llm
  kind: proxy
  routes:
    - when:
        - model:
            - gemini-*
      exit: gemini_llm_client

Next Steps