InstructorLite.Adapters.Gemini (instructor_lite v0.3.0)

View Source

Gemini adapter.

This adapter is implemented using Text generation endpoint configured for structured output

Params

params argument should be shaped as a models.GenerateBody request body

Example

InstructorLite.instruct(
  %{contents: [%{role: "user", parts: [%{text: "John is 25yo"}]}]},
  response_model: %{name: :string, age: :integer},
  json_schema: %{
    type: "object",
    required: [:age, :name],
    properties: %{name: %{type: "string"}, age: %{type: "integer"}}
  },
  adapter: InstructorLite.Adapters.Gemini,
  adapter_context: [
    model: "gemini-1.5-flash-8b",
    api_key: Application.fetch_env!(:instructor_lite, :gemini_key)
  ]
)
{:ok, %{name: "John", age: 25}}

Specifying model

Note how, unlike other adapters, the Gemini adapter expects model under adapter_context.

JSON Schema

Gemini's idea of JSON Schema is quite different from other major models, so InstructorLite.JSONSchema won't help you even for simple cases. Luckily, the Gemini API provides detailed errors for invalid schemas.

Summary

Functions

initial_prompt(params, opts)

Puts systemInstruction and updates generationConfig in params with prompt based on json_schema and notes.

parse_response(response, opts)

Parse text generation endpoint response.

retry_prompt(params, resp_params, errors, response, opts)

Updates params with prompt for retrying a request.

send_request(params, opts)

Make request to Gemini API

Functions

initial_prompt(params, opts)

Puts systemInstruction and updates generationConfig in params with prompt based on json_schema and notes.

parse_response(response, opts)

Parse text generation endpoint response.

Can return:

  • {:ok, parsed_json} on success.
  • {:error, :refusal, prompt_feedback} if request was blocked.
  • {:error, :unexpected_response, response} if response is of unexpected shape.

retry_prompt(params, resp_params, errors, response, opts)

Updates params with prompt for retrying a request.

send_request(params, opts)

Make request to Gemini API

Options

  • :api_key (String.t/0) - Required. Gemini API key

  • :http_client (atom/0) - Any module that follows Req.post/2 interface The default value is Req.

  • :http_options (keyword/0) - Options passed to http_client.post/2 The default value is [receive_timeout: 60000].

  • :url (String.t/0) - API endpoint to use for sending requests The default value is "https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent".

  • :model (String.t/0) - Gemini model The default value is "gemini-1.5-flash-8b".