Mistral Migration Deep Dive

Mistral AI Migration Deep Dive

Current State

!npm list openai @anthropic-ai/sdk @mistralai/mistralai 2>/dev/null | grep -E "openai|anthropic|mistral" || echo 'No AI SDKs found'

Overview

Comprehensive migration guide from OpenAI or Anthropic to Mistral AI using the adapter pattern with feature-flag controlled rollout. Covers model mapping, API differences, prompt adjustments, validation testing, and rollback procedures.

Prerequisites

• Current AI integration documented
• Mistral AI SDK installed (@mistralai/mistralai)
• Feature flag infrastructure (env vars or LaunchDarkly)
• Rollback plan tested

Migration Complexity

Instructions

Step 1: Assessment — Find All AI Touchpoints

set -euo pipefail
# Count integration points
echo "=== AI Integration Assessment ==="
echo "OpenAI imports: $(grep -r "from 'openai'" src/ --include='*.ts' -l 2>/dev/null | wc -l)"
echo "Anthropic imports: $(grep -r "from '@anthropic'" src/ --include='*.ts' -l 2>/dev/null | wc -l)"
echo "Chat completions: $(grep -r "chat\.completions\|messages\.create" src/ --include='*.ts' -c 2>/dev/null | wc -l)"
echo "Embeddings: $(grep -r "embeddings\.create" src/ --include='*.ts' -c 2>/dev/null | wc -l)"
echo "Streaming: $(grep -r "stream\|for await" src/ --include='*.ts' -c 2>/dev/null | wc -l)"

Step 2: Model Mapping

Step 3: Provider-Agnostic Adapter

// adapters/types.ts
export interface Message {
  role: 'system' | 'user' | 'assistant' | 'tool';
  content: string;
}

export interface ChatOptions {
  model?: string;
  temperature?: number;
  maxTokens?: number;
  stream?: boolean;
}

export interface ChatResponse {
  content: string;
  usage: { inputTokens: number; outputTokens: number };
  model: string;
}

export interface AIAdapter {
  chat(messages: Message[], options?: ChatOptions): Promise<ChatResponse>;
  chatStream(messages: Message[], options?: ChatOptions): AsyncGenerator<string>;
  embed(texts: string[]): Promise<number[][]>;
}

Step 4: Mistral Adapter

// adapters/mistral.adapter.ts
import { Mistral } from '@mistralai/mistralai';
import type { AIAdapter, Message, ChatOptions, ChatResponse } from './types.js';

export class MistralAdapter implements AIAdapter {
  private client: Mistral;
  private defaultModel: string;

  constructor(apiKey: string, defaultModel = 'mistral-small-latest') {
    this.client = new Mistral({ apiKey });
    this.defaultModel = defaultModel;
  }

  async chat(messages: Message[], options?: ChatOptions): Promise<ChatResponse> {
    const response = await this.client.chat.complete({
      model: options?.model ?? this.defaultModel,
      messages,
      temperature: options?.temperature,
      maxTokens: options?.maxTokens,
    });

    return {
      content: response.choices?.[0]?.message?.content ?? '',
      usage: {
        inputTokens: response.usage?.promptTokens ?? 0,
        outputTokens: response.usage?.completionTokens ?? 0,
      },
      model: response.model ?? this.defaultModel,
    };
  }

  async *chatStream(messages: Message[], options?: ChatOptions): AsyncGenerator<string> {
    const stream = await this.client.chat.stream({
      model: options?.model ?? this.defaultModel,
      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;
    }
  }

  async embed(texts: string[]): Promise<number[][]> {
    const response = await this.client.embeddings.create({
      model: 'mistral-embed',
      inputs: texts,
    });
    return response.data.map(d => d.embedding);
  }
}

Step 5: Feature-Flag Controlled Rollout

// adapters/factory.ts
import { MistralAdapter } from './mistral.adapter.js';
import { OpenAIAdapter } from './openai.adapter.js';

export function createAdapter(): AIAdapter {
  const rolloutPercent = parseInt(process.env.MISTRAL_ROLLOUT_PERCENT ?? '0');
  const useMistral = Math.random() * 100 < rolloutPercent;

  if (useMistral) {
    console.log('[AI] Using Mistral');
    return new MistralAdapter(process.env.MISTRAL_API_KEY!);
  }

  console.log('[AI] Using OpenAI (legacy)');
  return new OpenAIAdapter(process.env.OPENAI_API_KEY!);
}

Step 6: Gradual Rollout Plan

# Advance rollout
export MISTRAL_ROLLOUT_PERCENT=5   # Canary
export MISTRAL_ROLLOUT_PERCENT=25  # Partial
export MISTRAL_ROLLOUT_PERCENT=100 # Full migration
export MISTRAL_ROLLOUT_PERCENT=0   # Emergency rollback

Step 7: A/B Validation Testing

async function validateMigration(adapter1: AIAdapter, adapter2: AIAdapter) {
  const testPrompts = [
    'Summarize: TypeScript adds static typing to JavaScript.',
    'Classify: "The app crashes on login" — bug, feature, or question?',
    'What is 2+2?',
  ];

  for (const prompt of testPrompts) {
    const messages = [{ role: 'user' as const, content: prompt }];
    const [r1, r2] = await Promise.all([
      adapter1.chat(messages, { temperature: 0 }),
      adapter2.chat(messages, { temperature: 0 }),
    ]);

    console.log(`Prompt: ${prompt.slice(0, 50)}...`);
    console.log(`  Provider 1: ${r1.content.slice(0, 100)} (${r1.usage.outputTokens} tokens)`);
    console.log(`  Provider 2: ${r2.content.slice(0, 100)} (${r2.usage.outputTokens} tokens)`);
    console.log();
  }
}

Key API Differences

Error Handling

Examples

Migrate one classification endpoint safely

Put the old and Mistral adapters behind the same interface, send a fixed evaluation set through both at zero temperature, and compare category agreement, latency, and token cost. Start production at the five-percent flag, advance only when the agreed thresholds hold, and reset the flag to zero for an immediate rollback.

Resources

• Mistral AI Documentation
• Mistral vs OpenAI Comparison
• Strangler Fig Pattern

Output

• Integration assessment with effort estimation
• Provider-agnostic adapter interface
• Mistral adapter implementation
• Feature-flag controlled gradual rollout
• Model mapping and API difference reference
• A/B validation test suite
• Rollback procedure (set MISTRAL_ROLLOUT_PERCENT=0)