Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/mistral-reference-architecture/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill mistral-reference-architecture. Copyright stays with the author (MIT).

Mistral AI Reference Architecture

Overview

Production-ready architecture patterns for Mistral AI integrations: layered project structure, singleton client, Zod-validated config, custom error classes, service layer with caching, health checks, prompt templates, and model routing.

Prerequisites

• TypeScript/Node.js project (ESM)
• @mistralai/mistralai SDK
• zod for config validation
• Testing framework (Vitest)

Layer Architecture

API Layer (Routes, Controllers, Middleware)
    |
Service Layer (Business Logic, Orchestration)
    |
Mistral Layer (Client, Config, Errors, Prompts)
    |
Infrastructure Layer (Cache, Queue, Monitoring)

Instructions

Step 1: Directory Structure

src/
├── mistral/
│   ├── client.ts         # Singleton client factory
│   ├── config.ts         # Zod-validated config
│   ├── errors.ts         # Custom error classes
│   ├── types.ts          # Shared types
│   └── prompts.ts        # Prompt templates
├── services/
│   ├── chat.service.ts   # Chat with caching + retry
│   ├── embed.service.ts  # Embeddings + search
│   └── rag.service.ts    # RAG pipeline
├── api/
│   ├── chat.route.ts     # HTTP endpoints
│   └── health.route.ts   # Health check
└── config/
    ├── base.ts           # Shared config
    ├── development.ts    # Dev overrides
    └── production.ts     # Prod overrides

Step 2: Config with Zod Validation

// src/mistral/config.ts
import { z } from 'zod';

const MistralConfigSchema = z.object({
  apiKey: z.string().min(10, 'MISTRAL_API_KEY required'),
  defaultModel: z.string().default('mistral-small-latest'),
  timeoutMs: z.number().default(30_000),
  maxRetries: z.number().default(3),
  cache: z.object({
    enabled: z.boolean().default(true),
    ttlMs: z.number().default(3_600_000),
    maxSize: z.number().default(5000),
  }).default({}),
});

export type MistralConfig = z.infer<typeof MistralConfigSchema>;

export function loadConfig(): MistralConfig {
  return MistralConfigSchema.parse({
    apiKey: process.env.MISTRAL_API_KEY,
    defaultModel: process.env.MISTRAL_MODEL,
    timeoutMs: process.env.MISTRAL_TIMEOUT ? Number(process.env.MISTRAL_TIMEOUT) : undefined,
  });
}

Step 3: Singleton Client

// src/mistral/client.ts
import { Mistral } from '@mistralai/mistralai';
import { loadConfig, type MistralConfig } from './config.js';

let _client: Mistral | null = null;
let _config: MistralConfig | null = null;

export function getMistralClient(): Mistral {
  if (!_client) {
    _config = loadConfig();
    _client = new Mistral({
      apiKey: _config.apiKey,
      timeoutMs: _config.timeoutMs,
      maxRetries: _config.maxRetries,
    });
  }
  return _client;
}

export function getConfig(): MistralConfig {
  if (!_config) loadConfig();
  return _config!;
}

export function resetClient(): void {
  _client = null;
  _config = null;
}

Step 4: Custom Error Classes

// src/mistral/errors.ts
export type MistralErrorCode =
  | 'AUTH_ERROR'
  | 'RATE_LIMIT'
  | 'BAD_REQUEST'
  | 'SERVICE_ERROR'
  | 'TIMEOUT'
  | 'CONTEXT_OVERFLOW';

export class MistralServiceError extends Error {
  constructor(
    message: string,
    public readonly code: MistralErrorCode,
    public readonly status: number,
    public readonly retryable: boolean,
  ) {
    super(message);
    this.name = 'MistralServiceError';
  }

  static fromApiError(error: any): MistralServiceError {
    const status = error.status ?? error.statusCode ?? 500;
    if (status === 401) return new MistralServiceError('Authentication failed', 'AUTH_ERROR', 401, false);
    if (status === 429) return new MistralServiceError('Rate limit exceeded', 'RATE_LIMIT', 429, true);
    if (status === 400) return new MistralServiceError(error.message, 'BAD_REQUEST', 400, false);
    if (status >= 500) return new MistralServiceError('Service error', 'SERVICE_ERROR', status, true);
    return new MistralServiceError(error.message, 'SERVICE_ERROR', status, false);
  }
}

Step 5: Service Layer with Caching

// src/services/chat.service.ts
import { createHash } from 'crypto';
import { LRUCache } from 'lru-cache';
import { getMistralClient, getConfig } from '../mistral/client.js';
import { MistralServiceError } from '../mistral/errors.js';

const cache = new LRUCache<string, any>({ max: 5000, ttl: 3_600_000 });

export class ChatService {
  async complete(messages: any[], options?: {
    model?: string;
    temperature?: number;
    maxTokens?: number;
  }) {
    const config = getConfig();
    const model = options?.model ?? config.defaultModel;
    const temperature = options?.temperature ?? 0.7;

    // Cache deterministic requests
    if (temperature === 0 && config.cache.enabled) {
      const key = createHash('sha256').update(JSON.stringify({ model, messages })).digest('hex');
      const cached = cache.get(key);
      if (cached) return cached;

      const result = await this.executeChat(model, messages, { ...options, temperature: 0 });
      cache.set(key, result);
      return result;
    }

    return this.executeChat(model, messages, options);
  }

  async *stream(messages: any[], model?: string) {
    const client = getMistralClient();
    try {
      const stream = await client.chat.stream({
        model: model ?? getConfig().defaultModel,
        messages,
      });
      for await (const event of stream) {
        const text = event.data?.choices?.[0]?.delta?.content;
        if (text) yield text;
      }
    } catch (error: any) {
      throw MistralServiceError.fromApiError(error);
    }
  }

  private async executeChat(model: string, messages: any[], options: any = {}) {
    const client = getMistralClient();
    try {
      return await client.chat.complete({ model, messages, ...options });
    } catch (error: any) {
      throw MistralServiceError.fromApiError(error);
    }
  }
}

Step 6: Health Check

// src/api/health.route.ts
import { getMistralClient } from '../mistral/client.js';

export async function checkMistralHealth() {
  const start = performance.now();
  try {
    const client = getMistralClient();
    const models = await client.models.list();
    const latencyMs = Math.round(performance.now() - start);

    return {
      status: latencyMs > 5000 ? 'degraded' : 'healthy',
      latencyMs,
      modelCount: models.data?.length ?? 0,
    };
  } catch (error: any) {
    return {
      status: 'unhealthy',
      latencyMs: Math.round(performance.now() - start),
      error: error.message,
    };
  }
}

Step 7: Prompt Templates

// src/mistral/prompts.ts
interface PromptTemplate {
  name: string;
  system: string;
  userTemplate: (input: string) => string;
  model: string;
  maxTokens: number;
}

export const PROMPTS: Record<string, PromptTemplate> = {
  summarize: {
    name: 'summarize',
    system: 'Summarize the text in 2-3 sentences. Be factual and concise.',
    userTemplate: (text) => `Summarize:\n\n${text}`,
    model: 'mistral-small-latest',
    maxTokens: 200,
  },
  classify: {
    name: 'classify',
    system: 'Classify the input. Reply with one word only.',
    userTemplate: (text) => text,
    model: 'mistral-small-latest',
    maxTokens: 10,
  },
  codeReview: {
    name: 'codeReview',
    system: 'Review code for bugs, security issues, and improvements. Be specific.',
    userTemplate: (code) => `Review this code:\n\`\`\`\n${code}\n\`\`\``,
    model: 'mistral-large-latest',
    maxTokens: 1000,
  },
};

Error Handling

Examples

Wire a deterministic summarization service

Validate the environment at process start, route summarization through the singleton client with temperature zero and the bounded summarize template, and cache only the normalized input plus model version. On a retryable provider failure, surface a typed service error to the caller; on a 401, stop retrying and alert the credential owner.

Resources

• Mistral API Reference
• Clean Architecture
• 12-Factor App

Output

• Layered directory structure with clear separation
• Zod-validated configuration from environment
• Singleton client with lazy initialization
• Custom error classes with retryability
• Service layer with caching and streaming
• Health check endpoint
• Reusable prompt templates