Mistral AI SDK Patterns

Overview

Production-ready patterns for Mistral AI SDK usage in TypeScript and Python.

Prerequisites

• Completed mistral-install-auth setup
• Familiarity with async/await patterns
• Understanding of error handling best practices

Instructions

Step 1: Implement Singleton Pattern (Recommended)

TypeScript

// src/mistral/client.ts
import Mistral from '@mistralai/mistralai';

let instance: Mistral | null = null;

export interface MistralConfig {
  apiKey: string;
  timeout?: number;
  maxRetries?: number;
}

export function getMistralClient(config?: Partial<MistralConfig>): Mistral {
  if (!instance) {
    const apiKey = config?.apiKey || process.env.MISTRAL_API_KEY;
    if (!apiKey) {
      throw new Error('MISTRAL_API_KEY is required');
    }
    instance = new Mistral({
      apiKey,
      timeout: config?.timeout ?? 30000,
    });
  }
  return instance;
}

// For testing - reset singleton
export function resetMistralClient(): void {
  instance = null;
}

Python

# src/mistral/client.py
import os
from mistralai import Mistral
from typing import Optional

_instance: Optional[Mistral] = None

def get_mistral_client(api_key: Optional[str] = None) -> Mistral:
    global _instance
    if _instance is None:
        key = api_key or os.environ.get("MISTRAL_API_KEY")
        if not key:
            raise ValueError("MISTRAL_API_KEY is required")
        _instance = Mistral(api_key=key)
    return _instance

def reset_mistral_client() -> None:
    global _instance
    _instance = None

Step 2: Add Error Handling Wrapper

TypeScript

import Mistral from '@mistralai/mistralai';

interface MistralResult<T> {
  data: T | null;
  error: Error | null;
  usage?: { promptTokens: number; completionTokens: number; totalTokens: number };
}

async function safeMistralCall<T>(
  operation: () => Promise<T>
): Promise<MistralResult<T>> {
  try {
    const data = await operation();
    return { data, error: null };
  } catch (err) {
    const error = err as Error;
    console.error({
      name: error.name,
      message: error.message,
      // Extract status code if available
      status: (err as any).status,
    });
    return { data: null, error };
  }
}

// Usage
const result = await safeMistralCall(() =>
  client.chat.complete({
    model: 'mistral-small-latest',
    messages: [{ role: 'user', content: 'Hello!' }],
  })
);

if (result.error) {
  console.error('Chat failed:', result.error.message);
} else {
  console.log(result.data?.choices?.[0]?.message?.content);
}

Step 3: Implement Retry Logic with Exponential Backoff

interface RetryConfig {
  maxRetries: number;
  baseDelayMs: number;
  maxDelayMs: number;
  jitterMs: number;
}

async function withRetry<T>(
  operation: () => Promise<T>,
  config: RetryConfig = { maxRetries: 3, baseDelayMs: 1000, maxDelayMs: 32000, jitterMs: 500 }
): Promise<T> {
  for (let attempt = 0; attempt <= config.maxRetries; attempt++) {
    try {
      return await operation();
    } catch (err) {
      if (attempt === config.maxRetries) throw err;

      const status = (err as any).status;
      // Only retry on rate limits (429) or server errors (5xx)
      if (status !== 429 && (status < 500 || status >= 600)) throw err;

      const exponentialDelay = config.baseDelayMs * Math.pow(2, attempt);
      const jitter = Math.random() * config.jitterMs;
      const delay = Math.min(exponentialDelay + jitter, config.maxDelayMs);

      console.log(`Attempt ${attempt + 1} failed. Retrying in ${delay.toFixed(0)}ms...`);
      await new Promise(r => setTimeout(r, delay));
    }
  }
  throw new Error('Unreachable');
}

Step 4: Type-Safe Chat Wrapper

import Mistral from '@mistralai/mistralai';

type MistralModel =
  | 'mistral-large-latest'
  | 'mistral-medium-latest'
  | 'mistral-small-latest'
  | 'open-mistral-7b'
  | 'open-mixtral-8x7b';

interface ChatOptions {
  model?: MistralModel;
  temperature?: number;
  maxTokens?: number;
  topP?: number;
  stream?: boolean;
}

interface Message {
  role: 'system' | 'user' | 'assistant';
  content: string;
}

export async function chat(
  client: Mistral,
  messages: Message[],
  options: ChatOptions = {}
): Promise<string> {
  const response = await client.chat.complete({
    model: options.model ?? 'mistral-small-latest',
    messages,
    temperature: options.temperature,
    maxTokens: options.maxTokens,
    topP: options.topP,
  });

  return response.choices?.[0]?.message?.content ?? '';
}

Step 5: Streaming Helper

export async function* streamChat(
  client: Mistral,
  messages: Message[],
  options: ChatOptions = {}
): AsyncGenerator<string, void, unknown> {
  const stream = await client.chat.stream({
    model: options.model ?? 'mistral-small-latest',
    messages,
    temperature: options.temperature,
    maxTokens: options.maxTokens,
  });

  for await (const event of stream) {
    const content = event.data?.choices?.[0]?.delta?.content;
    if (content) {
      yield content;
    }
  }
}

// Usage
for await (const chunk of streamChat(client, messages)) {
  process.stdout.write(chunk);
}

Output

• Type-safe client singleton
• Robust error handling with structured logging
• Automatic retry with exponential backoff
• Streaming support with async generators

Error Handling

Examples

Factory Pattern (Multi-tenant)

const clients = new Map<string, Mistral>();

export function getClientForTenant(tenantId: string): Mistral {
  if (!clients.has(tenantId)) {
    const apiKey = getTenantApiKey(tenantId);
    clients.set(tenantId, new Mistral({ apiKey }));
  }
  return clients.get(tenantId)!;
}

Python Context Manager

from contextlib import contextmanager
from mistralai import Mistral

@contextmanager
def mistral_client():
    client = Mistral(api_key=os.environ.get("MISTRAL_API_KEY"))
    try:
        yield client
    finally:
        pass  # Cleanup if needed

# Usage
with mistral_client() as client:
    response = client.chat.complete(
        model="mistral-small-latest",
        messages=[{"role": "user", "content": "Hello!"}]
    )

Zod Response Validation

import { z } from 'zod';

const mistralResponseSchema = z.object({
  id: z.string(),
  object: z.literal('chat.completion'),
  model: z.string(),
  choices: z.array(z.object({
    index: z.number(),
    message: z.object({
      role: z.enum(['assistant']),
      content: z.string(),
    }),
    finishReason: z.string(),
  })),
  usage: z.object({
    promptTokens: z.number(),
    completionTokens: z.number(),
    totalTokens: z.number(),
  }),
});

function validateResponse(response: unknown) {
  return mistralResponseSchema.parse(response);
}

Resources

• Mistral AI API Reference
• Mistral AI Client Libraries
• Zod Documentation

Next Steps

Apply patterns in mistral-core-workflow-a for chat completions.