Mistral AI Reference Architecture

Table of Contents

• Overview
• Prerequisites
• Instructions
• Output
• Error Handling
• Examples
• Resources

Overview

Production-ready architecture patterns for Mistral AI integrations. Covers layered project structure, client singleton, configuration management, error handling, service layer with caching, health checks, and prompt templates.

Prerequisites

• Understanding of layered architecture
• Mistral AI SDK knowledge
• TypeScript project setup
• Testing framework configured

Layer Architecture

API Layer (Routes, Controllers, Middleware)
    ↓
Service Layer (Business Logic, Orchestration)
    ↓
Mistral Layer (Client, Prompts, Error Handling)
    ↓
Infrastructure Layer (Cache, Queue, Monitoring)

Instructions

Step 1: Create Directory Structure

Organize into src/mistral/ (client, config, types, errors, prompts, handlers), src/services/ai/ (chat, rag, cache), src/api/ (routes, streaming), tests/ (unit, integration), and config/ (per-environment JSON).

Step 2: Implement Client Singleton

Create getMistralClient() factory that lazily initializes a single Mistral instance with config from environment. Include resetMistralClient() for testing.

Step 3: Configure with Zod Validation

Define config schema with z.object() for apiKey, model (default mistral-small-latest), timeout (30s), maxRetries (3), and cache settings. Parse from environment variables.

Step 4: Build Error Handling Layer

Create MistralServiceError class with code, status, retryable flag. Wrap raw errors: 429 = RATE_LIMIT (retryable), 401 = AUTH_ERROR, 500+ = SERVICE_ERROR (retryable).

Step 5: Implement Service Layer

Build ChatService with complete() and stream() methods. Add cache check for deterministic requests (temperature=0). Wrap all calls with retry and error handling.

Step 6: Add Health Check

Implement checkMistralHealth() that calls client.models.list() and returns status (healthy/degraded/unhealthy) with latency measurement.

Step 7: Create Prompt Templates

Define reusable PromptTemplate interface with system prompt and user template function. Include templates for summarize, classify, and codeReview.

Output

• Structured project layout with clear layer separation
• Client singleton wrapper with retry logic
• Zod-validated configuration management
• Custom error classes with retryable detection
• Service layer with caching and streaming
• Health check endpoint
• Reusable prompt templates

Error Handling

Examples

Client Usage

const client = getMistralClient();
const response = await client.chat.complete({
  model: 'mistral-small-latest',
  messages: [{ role: 'user', content: 'Hello' }],
});

See detailed implementation for advanced patterns.

Resources

• Mistral AI API Reference
• Clean Architecture