openai

OpenAI Plugin

This plugin provides Genkit support for OpenAI chat models and embedders through the Chat Completions API.

Setup

Set an OpenAI API key:

export OPENAI_API_KEY=<your-api-key>

import (
    "context"

    "github.com/firebase/genkit/go/ai"
    "github.com/firebase/genkit/go/genkit"
    oai "github.com/firebase/genkit/go/plugins/compat_oai/openai"
)

ctx := context.Background()
plugin := &oai.OpenAI{}
g := genkit.Init(ctx,
    genkit.WithPlugins(plugin),
    genkit.WithDefaultModel("openai/gpt-5.4"),
)

response, err := genkit.Generate(ctx, g, ai.WithPrompt("Write a haiku about Go."))

Client options such as a different base URL or an organization ride the plugin's Opts (option.WithBaseURL, option.WithOrganization, and the rest of the SDK's request options).

Models

The plugin registers a curated catalog spanning the GPT-5 line (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, and earlier), the GPT-4 line, and the o-series reasoning models. The catalog is not a ceiling: any model ID the API serves resolves on demand, and the models the endpoint reports are listed dynamically. Use the Models field to describe or correct any model, most often one released after this plugin:

plugin := &oai.OpenAI{Models: map[string]ai.ModelOptions{
    "gpt-6": {Label: "GPT-6", Supports: &compat_oai.Multimodal},
}}

The current model list and capabilities are at https://developers.openai.com/api/docs/models.

Models served only by the Responses API (the -pro variants) are not part of this plugin, which speaks the Chat Completions API; naming one still resolves it and fails at request time.

Config

Models take the OpenAI SDK's own request type, openai.ChatCompletionNewParams, with the SDK's wire names. oai.ModelRef carries a config with the model ID:

import "github.com/openai/openai-go"

response, err := genkit.Generate(ctx, g,
    ai.WithModel(oai.ModelRef("gpt-5.4", &openai.ChatCompletionNewParams{
        Temperature:         openai.Float(0.2),
        MaxCompletionTokens: openai.Int(1024),
    })),
    ai.WithPrompt("Answer concisely."),
)

The GPT-5 generation rejects the legacy max_tokens, so cap output with MaxCompletionTokens; the older field remains for the models that still read it.

Set the config's Model to a dated snapshot to pin the exact version a request is served by. The fields Genkit builds from the request (messages, tools, and their variants) are not config: a config naming one is rejected.

The advertised schema is reflected from the openai-go version your build links, so a field OpenAI ships tomorrow becomes usable, and validated, by bumping github.com/openai/openai-go in your own go.mod.

Embedders

text-embedding-3-large, text-embedding-3-small, and text-embedding-ada-002 are registered, and the Embedders field overrides what the plugin knows the way Models does. Embedders take a typed oai.TextEmbeddingConfig:

res, err := genkit.Embed(ctx, g,
    ai.WithEmbedder(oai.NewEmbedderRef("text-embedding-3-small", &oai.TextEmbeddingConfig{
        Dimensions: 256,
    })),
    ai.WithTextDocs("Genkit is an AI framework."),
)

The embedding config also carries the settings Genkit owns: apiKey (settable only from Go code) serves one request with a different credential, and extra forwards request body fields the config does not declare, keyed by OpenAI's wire names.

Live tests

Live tests are skipped unless OPENAI_API_KEY is set:

go test -v ./plugins/compat_oai/openai