Azure OpenAI

Run text batches on Azure OpenAI Batch with Azure Global Batch or Data Zone Batch deployments, JSONL file upload, and normalized results.

Azure OpenAI Batch follows the OpenAI JSONL flow: Batchwork serializes requests through the AI SDK, uploads the input file, creates a 24-hour batch, polls it, and reads the output and error files. Use a Global Batch or Data Zone Batch deployment; standard Azure deployments cannot process Batch API jobs.

Example

import { azure } from "@ai-sdk/azure";
import { batch } from "batchwork";

const job = await batch({
  // Azure model ids are deployment names.
  model: azure("my-batch-deployment"),
  requests: [
    { customId: "fr", prompt: "Capital of France? One word." },
    { customId: "jp", prompt: "Capital of Japan? One word." },
  ],
});

azure("deployment") uses Azure’s Responses API. Use azure.chat("deployment") for Chat Completions. Each deployment must be configured for Global Batch or Data Zone Batch before submitting a job.

How it works

1. Build — each request becomes a JSONL line with its deployment name in body.model.
2. Upload — Batchwork uploads that JSONL to /openai/v1/files with purpose=batch.
3. Create — it creates a 24-hour batch through /openai/v1/batches.
4. Poll — Batchwork reads the batch status until it reaches a terminal state.
5. Read — it downloads output and error JSONL files into normalized BatchResults.

Azure input lines retain the AI SDK path, such as /v1/responses or /v1/chat/completions. Batchwork creates text batches with endpoint: "/v1/chat/completions".

Notes

Azure OpenAI support covers Batchwork’s text batch() API for Chat Completions and Responses models. batch.embeddings(), batch.images(), batch.transcriptions(), and batch.moderations() are not supported for Azure OpenAI. The deployment must use Global Batch or Data Zone Batch; a standard deployment cannot process Batch API jobs.

Azure accepts up to 100,000 requests in one 200 MB input file. Batchwork defaults to 50,000 requests for every provider, so raise that guardrail explicitly when needed:

await batch({
  model: azure("my-batch-deployment"),
  requests,
  limits: { maxRequests: 100_000 },
});

Azure measures batch capacity as enqueued tokens: the input tokens count against the deployment’s quota until the batch reaches a terminal state. The available quota varies by model, Global Batch vs. Data Zone Batch deployment, and subscription. Enable Azure’s dynamic quota where available to reduce submission failures, and consult Azure’s current Batch limits before sizing a workload.

Batchwork’s Azure adapter uploads directly through the Files API. That uses Azure’s 200 MB input-file limit and the default 500 input files per resource when files have no expiry. Azure can raise that file count with expires_after or use Bring Your Own Storage, but Batchwork does not expose either Azure-specific staging option yet. Output and error files do not count against those input-file limits.

Credentials

Set a resource name and API key:

AZURE_RESOURCE_NAME=my-azure-openai-resource
AZURE_API_KEY=…

AZURE_OPENAI_API_KEY is also accepted. Batchwork converts AZURE_RESOURCE_NAME to the Azure API root automatically. To use an explicit endpoint instead, pass either the resource’s /openai URL or its /openai/v1 API root:

await batch({
  model: azure.chat("my-batch-deployment"),
  baseURL: "https://my-azure-openai-resource.openai.azure.com/openai/v1",
  apiKey: process.env.AZURE_API_KEY,
  requests,
});

Batchwork sends API keys as Azure’s api-key header. For Microsoft Entra ID, omit apiKey and pass your own bearer token in headers; caller-supplied Authorization or api-key headers take precedence over environment credentials. (An explicit apiKey argument always adds the api-key header, so don’t combine it with a bearer token.)

await batch({
  model: azure("my-batch-deployment"),
  baseURL: "https://my-azure-openai-resource.openai.azure.com/openai/v1",
  headers: { Authorization: `Bearer ${token}` },
  requests,
});

Install the optional peer dependency:

bun add @ai-sdk/azure