Setup

Install the Mistral adapter beside Anvia core:

pnpm add @anvia/core @anvia/mistral

Set MISTRAL_API_KEY in the server environment and create the client in a server-only module:

import { MistralClient } from '@anvia/mistral'

export const mistral = new MistralClient({
  apiKey: process.env.MISTRAL_API_KEY!,
})

The constructor throws when it must create an SDK client and the API key is missing or empty. Keep credentials and provider clients out of browser bundles; browser requests should pass through your application server.

Configure the SDK endpoint

MistralClient accepts three constructor options:

• apiKey authenticates the SDK client created by the adapter.
• baseUrl points that SDK client at an alternate Mistral-compatible endpoint.
• client reuses an existing Mistral SDK instance.

For example, keep an alternate server URL in trusted deployment configuration:

const mistral = new MistralClient({
  apiKey: process.env.MISTRAL_API_KEY!,
  baseUrl: process.env.MISTRAL_BASE_URL,
})

An alternate endpoint may support basic completions but differ for streaming, tools, schemas, embeddings, OCR, or model listing. Smoke-test every capability the application uses.

Bring an existing SDK client

Install the Mistral SDK directly when another infrastructure module constructs it:

pnpm add @mistralai/mistralai

import { Mistral } from '@mistralai/mistralai'
import { MistralClient } from '@anvia/mistral'

const sdk = new Mistral({
  apiKey: process.env.MISTRAL_API_KEY!,
})

export const mistral = new MistralClient({ client: sdk })

When client is supplied, the adapter uses it directly. This is useful for dependency injection or shared SDK transport configuration.

Export configured models

Create named model dependencies at the application boundary:

export const supportModel = mistral.completionModel({
    modelId: 'mistral-large-latest'
})

export const knowledgeEmbeddings = mistral.embeddingModel({
    modelId: 'mistral-embed'
})

export const documentOcr = mistral.ocrModel({
    modelId: 'mistral-ocr-latest'
})

Known IDs provide editor autocomplete, while custom string IDs remain valid. Treat each selected ID as deployment configuration and verify it against the target account and endpoint.