Mistral provider

@anvia/mistral adapts the Mistral SDK to Anvia completion, embedding, OCR, and model-listing contracts. Use it for Mistral chat workflows, retrieval embeddings, document OCR, and operational model inventory.

Install

pnpm add @anvia/core @anvia/mistral

Create the client in server-only code:

import { MistralClient } from "@anvia/mistral";

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

MistralClient accepts apiKey, serverURL, or an already-created Mistral SDK client.

Completion Models

import { AgentBuilder } from "@anvia/core";
import { MistralClient } from "@anvia/mistral";

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

const model = mistral.completionModel("mistral-large-latest");

export const agent = new AgentBuilder("support", model)
  .instructions("Answer support questions clearly.")
  .build();

MistralCompletionModel supports streaming, tools, tool choice, and output schemas at the Anvia contract level. It does not declare chat image input, chat document input, or reasoning support.

Server URL

Use serverURL when the Mistral SDK should target a configured server URL:

const mistral = new MistralClient({
  apiKey: process.env.MISTRAL_API_KEY,
  serverURL: "https://mistral.example.com",
});

Alternate server URLs should be smoke tested with tools, schemas, and streaming before production use.

Embeddings

const embeddings = mistral.embeddingModel("mistral-embed", {
  dimensions: 1024,
  maxBatchSize: 512,
});

const vectors = await embeddings.embedTexts([
  "Refunds take five business days.",
]);

MistralEmbeddingModel validates provider response indexes and preserves input order.

OCR

const ocr = mistral.ocrModel();

const result = await ocr.ocr({
  source: {
    type: "document_url",
    url: "https://example.com/invoice.pdf",
    documentName: "invoice.pdf",
  },
  includeImageBase64: false,
});

console.log(result.markdown);

OCR sources can be document URLs, image URLs, existing file ids, or bytes. Byte sources are uploaded through the Mistral files API before OCR processing.

The OCR response includes combined text and markdown, normalized page entries, optional document annotations, optional uploaded file metadata, usage information when returned, and the raw provider response.

Model Listing

const models = await mistral.listModels();

Use listing for inventory. Keep a separate app allowlist for production-enabled model ids.

Unsupported Features

The current adapter does not implement chat image input, chat document file input, transcription, audio generation, or image generation. Use Capability matrix before choosing Mistral for media workflows.

Exports

The root package exports MistralClient, MistralCompletionModel, MistralEmbeddingModel, MistralOcrModel, OCR request and response types, MISTRAL_OCR_LATEST, Mistral message conversion helpers, and the mistral namespace.