OpenAI media models

OpenAI image generation, speech generation, and transcription use separate Anvia model contracts.

1. Create capability-specific models

import {
  GPT_IMAGE_1,
  TTS_1,
  WHISPER_1,
  OpenAIClient,
} from '@anvia/openai'

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

const imageModel =
  openai.imageGenerationModel({
      modelId: GPT_IMAGE_1
  })
const audioModel =
  openai.speechGenerationModel({
      modelId: TTS_1
  })
const transcriptionModel =
  openai.transcriptionModel({
      modelId: WHISPER_1
  })

The package also exports known constants such as GPT_IMAGE_2, DALL_E_2, DALL_E_3, and TTS_1_HD. Model-name types accept custom IDs, but the endpoint must implement them.

2. Generate image bytes

import { generateImage } from '@anvia/core/image-generation'

const image = await generateImage({
    prompt: 'A minimal isometric illustration of a support workflow',
    model: imageModel,
    width: 1200,
    height: 800,
    retries: { maxAttempts: 3 }
})

await mediaStore.put({
  bytes: image.images[0].data,
  mediaType: image.images[0].mediaType ?? 'image/png',
})

The adapter maps base64 output to Uint8Array. It rejects URL-only image responses because the Core image response contract requires bytes.

The adapter applies its normalized fields (model, prompt, size) after providerOptions, so they win on conflict; provider keys the adapter never sets pass through unchanged. Keep providerOptions in trusted allow-listed configuration.

3. Generate speech

import { generateSpeech } from '@anvia/core/speech-generation'

const speech = await generateSpeech({
    text: 'Your incident report is ready.',
    model: audioModel,
    voice: 'alloy',
    speed: 1
})

The adapter returns audio bytes and a media type inferred from the requested response format. Validate script length, voice, speed, and format before the provider call.

4. Transcribe audio

import { transcribe } from '@anvia/core/transcription'

const transcript = await transcribe({
    audio: {
        data: upload.bytes,
        filename: upload.filename
    },
    model: transcriptionModel,
    language: 'en',
    retries: { maxAttempts: 3 }
})

console.log(transcript.text)

A meaningful filename helps the provider infer the media type. Validate ownership, type, and size before loading audio.

5. Store media outside runtime state

Persist media in object storage and pass asset references through application state. Keep raw bytes out of memory, traces, and event payloads. Apply access and retention policy to transcripts and generated assets.

See Multimodal workflows for composition guidance.