@inferagraph/azure-foundry-provider

Azure AI Foundry provider plugin for @inferagraph/core.

Supports any model in the Azure AI model catalog via the Azure AI Inference SDK (@azure-rest/ai-inference).

Installation

pnpm add @inferagraph/azure-foundry-provider @inferagraph/core

Usage

import { AzureFoundryProvider } from '@inferagraph/azure-foundry-provider';

const provider = new AzureFoundryProvider({
  endpoint: 'https://your-endpoint.inference.ai.azure.com',
  apiKey: 'your-api-key',
  deploymentName: 'gpt-4o',  // optional, sent as `model` on the request
  maxTokens: 1024,           // optional, default 1024
});

Configuration

| Option | Required | Description | |---|---|---| | endpoint | Yes | Azure AI Foundry endpoint URL | | apiKey | One of apiKey / credential | Azure key for the deployment | | credential | One of apiKey / credential | Any @azure/core-auth TokenCredential (e.g., DefaultAzureCredential) | | deploymentName | No | Sent as model in the request body when set | | maxTokens | No | Default max_tokens (1024) |

import { DefaultAzureCredential } from '@azure/identity';

new AzureFoundryProvider({
  endpoint: 'https://your-endpoint.inference.ai.azure.com',
  credential: new DefaultAzureCredential(),
});

Capabilities

The provider implements @inferagraph/core's LLMProvider contract:

| Method | Supported | Notes | |---|---|---| | complete(prompt, opts?) | Yes | Single-shot completion via /chat/completions. | | stream(prompt, opts?) | Yes | Single-string streaming. Kept for back-compat — new consumers should prefer streamMessages. | | streamMessages(messages, opts?) | Yes | Structured [{role, content}] streaming. system / user / assistant roles map 1:1 onto Foundry's OpenAI-compatible messages array, so system instructions stay separate from user input end-to-end. Honors opts.signal (AbortController), opts.maxTokens, opts.temperature, and opts.tools. | | embed(texts, opts?) | No | The @azure-rest/ai-inference SDK targets chat-completion routes only; this provider has no native embedding endpoint, so embed is intentionally omitted ('embed' in provider === false). Mirrors the @inferagraph/anthropic-provider no-Voyage path. Hosts that need embeddings pair Foundry chat with a separate embedding-capable provider (for example @inferagraph/openai-provider configured with an embeddingDeployment). The structural absence lets AIEngine detect the missing capability and route embedding work elsewhere. |

import { AzureFoundryProvider } from '@inferagraph/azure-foundry-provider';

const provider = new AzureFoundryProvider({
  endpoint: 'https://your-endpoint.inference.ai.azure.com',
  apiKey: 'your-api-key',
});

for await (const event of provider.streamMessages([
  { role: 'system', content: 'You are a helpful assistant.' },
  { role: 'user', content: 'Hi.' },
])) {
  if (event.type === 'text') process.stdout.write(event.delta);
}

License

MIT