`@anvia/openai`

OpenAI’s provider adapter covers the broadest set of Anvia model contracts: completions, embeddings, image generation, speech generation, transcription, and model listing. Use it at the model boundary while keeping agents and application logic provider-independent.


Support	First-party
Version	`1.1.1`
Runtime	ESM, server-side JavaScript
Peer	Matching `@anvia/core` stable release

Install

bash

pnpm add @anvia/openai @anvia/core

Create a completion model

import { Agent } from '@anvia/core'
import { OpenAIClient } from '@anvia/openai'

const openai = new OpenAIClient({
  apiKey: process.env.OPENAI_API_KEY!,
})

const agent = new Agent({
  id: 'assistant',
  model: openai.completionModel({
      modelId: 'gpt-5.6-sol',
      api: "responses"
  }),
  instructions: 'Answer clearly and concisely.',
})

const result = await agent.generate({
    prompt: 'Explain semantic search in one paragraph.'
})

if (result.type === 'response') {
  console.log(result.output)
}

The required api model option selects OpenAI Responses or Chat Completions. A custom baseUrl changes the endpoint but does not select the API.

Capabilities

Capability	Factory	Default
Streaming completion	`completionModel({ modelId, api })`	Explicit model and API
Dense embeddings	`embeddingModel({ modelId })`	Explicit model
Image generation	`imageGenerationModel({ modelId })`	Explicit model
Text-to-speech	`speechGenerationModel({ modelId })`	Explicit model
Transcription	`transcriptionModel({ modelId })`	Explicit model
Model inventory	`listModels()`	Provider model list

Both completion adapters normalize messages, tool calls, reasoning, usage, structured output, and streaming events into Anvia contracts. The Responses adapter is the native OpenAI path; the Chat adapter supports OpenAI-compatible endpoints and preserves provider-specific reasoning history when required. Reasoning effort is a typed control: completionModel({ modelId, api, controls }) accepts a per-model effort override, and individual requests can pass controls on top of the model defaults (see Configuration).

Common patterns

Configure an OpenAI-compatible endpoint

const compatible = new OpenAIClient({
  apiKey: process.env.PROVIDER_API_KEY!,
  baseUrl: 'https://provider.example.com/v1',
})

const model = compatible.completionModel({
    modelId: 'provider/model-name',
    api: 'chat',
})

Reuse one client across model capabilities

const completion = openai.completionModel({
    modelId: 'gpt-5.6-sol',
    api: 'responses',
})
const embeddings = openai.embeddingModel({
    modelId: 'text-embedding-3-small',
    dimensions: 1536,
    maxBatchSize: 64
})
const images = openai.imageGenerationModel({ modelId: 'gpt-image-2' })

Create the provider client once at the server boundary. Keep credentials there, inject the returned model contracts, and keep fallback policy and tenant routing in application code.

Compatibility

@anvia/openai is an ESM package and uses the official openai SDK. It accepts a preconfigured SDK client for custom transports. Media and embedding methods require endpoints that implement the corresponding OpenAI APIs; an OpenAI-compatible chat endpoint does not imply support for those other capabilities.

@anvia/openai ​

Install ​

Create a completion model ​

Capabilities ​

Common patterns ​

Configure an OpenAI-compatible endpoint ​

Reuse one client across model capabilities ​

Compatibility ​

Continue ​