Azure OpenAI Provider

Overview

The Azure OpenAI provider connects core-ai to Azure OpenAI deployments through the official OpenAI SDK. Azure v1 uses the Responses API for chatModel() and exposes strict Chat Completions through azure.chat.chatModel(). Pass Azure OpenAI deployment names as model IDs.

Installation

npm install @core-ai/azure-openai

createAzureOpenAI()

Create an Azure OpenAI provider instance.

import { createAzureOpenAI } from '@core-ai/azure-openai';

const azure = createAzureOpenAI({
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
});

Options

string

Your Azure OpenAI API key.

string

Your Azure OpenAI resource endpoint, for example https://my-resource.openai.azure.com. You can also pass a full v1 base URL such as https://my-resource.openai.azure.com/openai/v1.

'v1' | 'classic'

API surface to use. Defaults to v1. Set classic to use Azure’s classic API surface.

string

Required when api is classic. Azure OpenAI API version, for example 2025-04-01-preview.

string

Classic mode only. Default Azure OpenAI deployment for the underlying SDK client.

() => Promise<string>

Classic mode only. Token provider for Microsoft Entra ID authentication.

OpenAI | OpenAIChatClient

Provide your own configured client. Azure v1 clients must expose Responses and Chat Completions. Classic clients only need Chat Completions.

Returns

AzureOpenAIProvider with chatModel() and chat.chatModel().

API selection

With Azure v1, the root provider uses Responses:

const responsesModel = azure.chatModel('gpt-5-mini-deployment');

Use the nested provider for strict Chat Completions:

const chatCompletionsModel = azure.chat.chatModel('gpt-5-mini-deployment');

Classic Azure supports Chat Completions only. In classic mode, both azure.chatModel() and azure.chat.chatModel() create Chat Completions models.

Authentication

Use an API key with the v1 API:

const azure = createAzureOpenAI({
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
});

Use the classic API by setting api: 'classic' and passing an API version:

const azure = createAzureOpenAI({
    api: 'classic',
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
    apiVersion: '2025-04-01-preview',
});

Classic mode can use Microsoft Entra ID by passing a token provider:

import {
    DefaultAzureCredential,
    getBearerTokenProvider,
} from '@azure/identity';
import { createAzureOpenAI } from '@core-ai/azure-openai';

const credential = new DefaultAzureCredential();
const azureADTokenProvider = getBearerTokenProvider(
    credential,
    'https://cognitiveservices.azure.com/.default'
);

const azure = createAzureOpenAI({
    api: 'classic',
    azureADTokenProvider,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
    apiVersion: '2025-04-01-preview',
});

Model IDs

Use your Azure OpenAI deployment name as the model id:

const model = azure.chatModel('customer-support-gpt-5-mini');

If your deployment name is different from the base model name, pass the deployment name.

Model capabilities

Azure models expose the shared OpenAI capability shape through model.capabilities:

const model = azure.chatModel('gpt-5-mini');
const capabilities = model.capabilities;

Capability lookup uses the model ID passed to chatModel(). Prefer a known OpenAI model ID when that matches your Azure deployment name, so reasoning and image-input flags reflect the underlying model. If the value is a custom deployment name that is not in the OpenAI registry, the provider cannot infer the model family and returns the generic OpenAI fallback. Treat that data as a default rather than an exact description of the deployment. The fallback includes 'image' and 'file' in modalities.input, so multimodal requests are forwarded to the deployment rather than rejected locally. For recognized OpenAI model IDs, see OpenAI model capabilities for the reasoning, image input, and Chat Completions fields.

Reasoning metadata

On Azure v1 Responses, encrypted reasoning is stored under the azure-openai provider key — not openai. Use that key when reading metadata, and keep escape-hatch request options under providerOptions.openai (shared Responses field shapes).

import { generate, getProviderMetadata } from '@core-ai/core-ai';
import type { OpenAIReasoningMetadata } from '@core-ai/openai';
import { createAzureOpenAI } from '@core-ai/azure-openai';

const azure = createAzureOpenAI({
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
});

const result = await generate({
    model: azure.chatModel('gpt-5.4'),
    messages: [{ role: 'user', content: 'Think carefully before answering.' }],
    reasoning: { effort: 'high' },
});

for (const part of result.parts) {
    if (part.type !== 'reasoning') continue;

    const metadata = getProviderMetadata<OpenAIReasoningMetadata>(
        part.providerMetadata,
        'azure-openai'
    );

    console.log(metadata?.encryptedContent);
}

Examples

Basic chat

import { generate } from '@core-ai/core-ai';
import { createAzureOpenAI } from '@core-ai/azure-openai';

const azure = createAzureOpenAI({
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
});

const result = await generate({
    model: azure.chatModel('gpt-5-mini-deployment'),
    messages: [{ role: 'user', content: 'Hello!' }],
});

console.log(result.content);

Streaming

import { stream } from '@core-ai/core-ai';
import { createAzureOpenAI } from '@core-ai/azure-openai';

const azure = createAzureOpenAI({
    apiKey: process.env.AZURE_OPENAI_API_KEY,
    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
});

const result = await stream({
    model: azure.chatModel('gpt-5-mini-deployment'),
    messages: [{ role: 'user', content: 'Count to five.' }],
});

for await (const chunk of result.textStream) {
    process.stdout.write(chunk);
}

Provider-specific options

Root Azure v1 models use the OpenAI Responses options:

await generate({
    model: azure.chatModel('gpt-5-mini-deployment'),
    messages: [{ role: 'user', content: 'Hello!' }],
    providerOptions: {
        openai: {
            store: false,
            serviceTier: 'auto',
        },
    },
});

Nested and classic models use Chat Completions options:

await generate({
    model: azure.chat.chatModel('gpt-5-mini-deployment'),
    messages: [{ role: 'user', content: 'Hello!' }],
    providerOptions: {
        openai: {
            seed: 42,
        },
    },
});

See the OpenAI provider for the fields supported by each API.

Azure Responses availability depends on the model and Azure region. Use azure.chat.chatModel() when your deployment only supports Chat Completions.

Core

Providers

Azure OpenAI Provider

Overview

Installation

createAzureOpenAI()

Options

Returns

API selection

Authentication

Model IDs

Model capabilities

Reasoning metadata

Examples

Basic chat

Streaming

Provider-specific options

​Overview

​Installation

​createAzureOpenAI()

​Options

​Returns

​API selection

​Authentication

​Model IDs

​Model capabilities

​Reasoning metadata

​Examples

​Basic chat

​Streaming

​Provider-specific options

Overview

Installation

createAzureOpenAI()

Options

Returns

API selection

Authentication

Model IDs

Model capabilities

Reasoning metadata

Examples

Basic chat

Streaming

Provider-specific options