Mistral AI Core Workflow A: Chat Completions

Overview

Production chat completion patterns for Mistral AI: multi-turn conversations, streaming responses, JSON mode structured output, guardrails/moderation, and model selection. Uses the @mistralai/mistralai SDK.

Prerequisites

• Completed mistral-install-auth setup
• MISTRAL_API_KEY environment variable set
• Understanding of Mistral model tiers

Instructions

Step 1: Basic Chat Completion

import { Mistral } from '@mistralai/mistralai';

const client = new Mistral({ apiKey: process.env.MISTRAL_API_KEY });

async function chat(userMessage: string): Promise<string> {
  const response = await client.chat.complete({
    model: 'mistral-small-latest',
    messages: [
      { role: 'system', content: 'You are a helpful assistant.' },
      { role: 'user', content: userMessage },
    ],
  });
  return response.choices?.[0]?.message?.content ?? '';
}

Step 2: Multi-Turn Conversation Manager

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

class MistralConversation {
  private messages: Message[] = [];
  private client: Mistral;
  private model: string;

  constructor(systemPrompt: string, model = 'mistral-small-latest') {
    this.client = new Mistral({ apiKey: process.env.MISTRAL_API_KEY });
    this.model = model;
    this.messages.push({ role: 'system', content: systemPrompt });
  }

  async send(userMessage: string): Promise<string> {
    this.messages.push({ role: 'user', content: userMessage });

    const response = await this.client.chat.complete({
      model: this.model,
      messages: this.messages,
    });

    const reply = response.choices?.[0]?.message?.content ?? '';
    this.messages.push({ role: 'assistant', content: reply });
    return reply;
  }

  // Prevent context window overflow
  trimHistory(maxTurns = 20): void {
    const system = this.messages[0];
    const recent = this.messages.slice(1).slice(-maxTurns * 2);
    this.messages = [system, ...recent];
  }
}

// Usage
const conv = new MistralConversation('You are a coding tutor.');
await conv.send('How do I reverse a list in Python?');
await conv.send('What about in-place?');

Step 3: Streaming Responses

async function streamChat(
  messages: Message[],
  onChunk: (text: string) => void,
): Promise<string> {
  const stream = await client.chat.stream({
    model: 'mistral-small-latest',
    messages,
  });

  let full = '';
  for await (const event of stream) {
    const text = event.data?.choices?.[0]?.delta?.content;
    if (text) {
      full += text;
      onChunk(text);
    }
  }
  return full;
}

// Express.js SSE endpoint
app.post('/chat/stream', async (req, res) => {
  res.setHeader('Content-Type', 'text/event-stream');
  res.setHeader('Cache-Control', 'no-cache');
  res.setHeader('Connection', 'keep-alive');

  const stream = await client.chat.stream({
    model: 'mistral-small-latest',
    messages: req.body.messages,
  });

  for await (const event of stream) {
    const content = event.data?.choices?.[0]?.delta?.content;
    if (content) {
      res.write(`data: ${JSON.stringify({ content })}\n\n`);
    }
  }
  res.write('data: [DONE]\n\n');
  res.end();
});

Step 4: JSON Mode and JSON Schema Mode

// JSON mode — model returns valid JSON
const jsonResponse = await client.chat.complete({
  model: 'mistral-small-latest',
  messages: [
    { role: 'user', content: 'List 3 countries with capitals as JSON array.' },
  ],
  responseFormat: { type: 'json_object' },
});
const data = JSON.parse(jsonResponse.choices?.[0]?.message?.content ?? '{}');

// JSON Schema mode — guarantees structure conformance
const schemaResponse = await client.chat.complete({
  model: 'mistral-small-latest',
  messages: [
    { role: 'user', content: 'Classify this ticket: "Login page crashes on mobile"' },
  ],
  responseFormat: {
    type: 'json_schema',
    jsonSchema: {
      name: 'ticket_classification',
      schema: {
        type: 'object',
        properties: {
          category: { type: 'string', enum: ['bug', 'feature', 'question'] },
          severity: { type: 'string', enum: ['low', 'medium', 'high', 'critical'] },
          summary: { type: 'string' },
        },
        required: ['category', 'severity', 'summary'],
      },
    },
  },
});

Step 5: Guardrails and Moderation

// Built-in safe_prompt flag — injects safety system prompt
const safeResponse = await client.chat.complete({
  model: 'mistral-small-latest',
  messages: [{ role: 'user', content: userInput }],
  safePrompt: true,
});

// Dedicated moderation API — classify text against policy categories
const moderation = await client.classifiers.moderate({
  model: 'mistral-moderation-latest',
  inputs: [userInput],
});

const flagged = moderation.results[0].categories;
// Check: flagged.sexual, flagged.hate_and_discrimination, flagged.violence, etc.
if (Object.values(flagged).some(Boolean)) {
  throw new Error('Content flagged by moderation');
}

Step 6: Model Selection Guide

type UseCase = 'realtime' | 'analysis' | 'code' | 'vision' | 'embedding';

const MODEL_MAP: Record<UseCase, { model: string; note: string }> = {
  realtime:  { model: 'mistral-small-latest',   note: '256k ctx, fast, $0.1/M in' },
  analysis:  { model: 'mistral-large-latest',   note: '256k ctx, reasoning, $0.5/M in' },
  code:      { model: 'codestral-latest',        note: '256k ctx, code + FIM, $0.3/M in' },
  vision:    { model: 'pixtral-large-latest',    note: '128k ctx, multimodal' },
  embedding: { model: 'mistral-embed',           note: '1024-dim vectors, $0.1/M in' },
};

function selectModel(use: UseCase): string {
  return MODEL_MAP[use].model;
}

Output

• Chat completions with configurable parameters
• Multi-turn conversation management with history trimming
• Real-time streaming responses
• JSON and JSON Schema structured output
• Content moderation via guardrails

Error Handling

Resources

• Chat Completions API
• JSON Mode
• Guardrails
• Models Overview

Next Steps

For embeddings and function calling, see mistral-core-workflow-b.