@mayurajs/provider-mistral

Mistral models for Mayura, through the official @mistralai/mistralai SDK and the Chat Completions API.

npm install mayura @mayurajs/provider-mistral

import { createModels, defineAgent, z } from 'mayura';
import { mistral } from '@mayurajs/provider-mistral';

const models = createModels({
  providers: [mistral({ apiKey: process.env.MISTRAL_API_KEY! })],
  prices: { 'mistral/mistral-medium-3-5': { inputMicrosPerMillionTokens: 1_500_000, outputMicrosPerMillionTokens: 7_500_000 } },
  maxCallCostMicros: 50_000,
});

const agent = defineAgent({
  id: 'helper', version: '1', instructions: 'Answer in one sentence.', model: models.model('mistral/mistral-medium-3-5'), tools: [],
  input: z.object({ question: z.string() }), output: z.object({ answer: z.string() }),
});
// Grant it as model:mistral/mistral-medium-3-5.

• Credentials are options: nothing is read from the environment. Set baseURL: 'https://api.eu.mistral.ai' to keep requests in the EU.
• The SDK's own retries, debug logging and telemetry are off; the registry's retry option retries and charges every attempt.
• Tool inputs and structured output use strict JSON Schema; streaming reports output text as it arrives; the model's thinking is kept for the next call and never released.
• A run's conversation history (capabilities.history) goes first, as Mistral messages; a call id Mistral would refuse is renamed, never to one the conversation already has.
• Models see PNG, JPEG, WebP and GIF images, as bytes or URLs.
• catalog holds Mistral's list prices on its date for dated model versions, used only with prices: 'catalog'. A -latest alias moves to newer models, so give its price yourself.

See the model registry guide. Apache-2.0.

RAG embeddings

This provider exports mistralEmbeddings for scoped RAG ingestion and retrieval. See the RAG guide for explicit permissions, cost ceilings and sensitivity profiles. Development cases are authored; validation is deferred.