Mistral CI Integration

Mistral CI Integration

Overview

Integrate Mistral AI validation into CI/CD pipelines: prompt regression tests, model response quality checks, cost estimation in PR comments, and deployment gates for prompt changes. Uses GitHub Actions with MISTRAL_API_KEY stored as a repository secret.

Prerequisites

• MISTRAL_API_KEY stored as GitHub repository secret
• GitHub Actions configured
• Test framework (Vitest recommended)

Instructions

Step 1: GitHub Actions Workflow

# .github/workflows/mistral-tests.yml
name: Mistral AI Tests

on:
  pull_request:
    paths:
      - 'src/prompts/**'
      - 'src/ai/**'
      - 'tests/ai/**'

jobs:
  prompt-tests:
    runs-on: ubuntu-latest
    timeout-minutes: 10
    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-node@v4
        with:
          node-version: '20'
          cache: 'npm'

      - run: npm ci

      - name: Run prompt regression tests
        env:
          MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
        run: npx vitest run tests/ai/ --reporter=verbose

      - name: Cost estimation
        env:
          MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
        run: npx tsx scripts/estimate-costs.ts >> $GITHUB_STEP_SUMMARY

Step 2: Prompt Regression Tests

// tests/ai/mistral-prompts.test.ts
import { describe, it, expect } from 'vitest';
import { Mistral } from '@mistralai/mistralai';

const apiKey = process.env.MISTRAL_API_KEY;

describe.skipIf(!apiKey)('Mistral Prompt Regression', () => {
  const client = new Mistral({ apiKey: apiKey! });

  it('summarization produces 2-3 sentences', async () => {
    const result = await client.chat.complete({
      model: 'mistral-small-latest',
      messages: [
        { role: 'system', content: 'Summarize in 2-3 sentences.' },
        { role: 'user', content: 'TypeScript is a typed superset of JavaScript that compiles to plain JavaScript. It adds optional static typing and class-based OOP.' },
      ],
      maxTokens: 150,
      temperature: 0, // Deterministic for regression
    });

    const content = result.choices?.[0]?.message?.content ?? '';
    expect(content.length).toBeGreaterThan(20);
    expect(content.split(/[.!?]+/).filter(Boolean).length).toBeGreaterThanOrEqual(2);
  }, 15_000);

  it('classification returns valid category', async () => {
    const result = await client.chat.complete({
      model: 'mistral-small-latest',
      messages: [
        { role: 'system', content: 'Classify as: bug, feature, question. Reply with one word only.' },
        { role: 'user', content: 'The login page crashes on mobile devices' },
      ],
      maxTokens: 10,
      temperature: 0,
    });

    const category = result.choices?.[0]?.message?.content?.trim().toLowerCase();
    expect(['bug', 'feature', 'question']).toContain(category);
  }, 15_000);

  it('JSON mode returns valid JSON', async () => {
    const result = await client.chat.complete({
      model: 'mistral-small-latest',
      messages: [{ role: 'user', content: 'Return {"status": "ok"} as JSON' }],
      responseFormat: { type: 'json_object' },
      maxTokens: 50,
      temperature: 0,
    });

    const content = result.choices?.[0]?.message?.content ?? '';
    expect(() => JSON.parse(content)).not.toThrow();
  }, 15_000);

  it('respects maxTokens limit', async () => {
    const result = await client.chat.complete({
      model: 'mistral-small-latest',
      messages: [{ role: 'user', content: 'Write a long essay about AI' }],
      maxTokens: 50,
    });

    expect(result.usage?.completionTokens).toBeLessThanOrEqual(50);
  }, 15_000);
});

Step 3: Cost Estimation Script

// scripts/estimate-costs.ts
import { readdirSync, readFileSync } from 'fs';
import { join } from 'path';

const PRICING: Record<string, { input: number; output: number }> = {
  'mistral-small-latest': { input: 0.1, output: 0.3 },
  'mistral-large-latest': { input: 0.5, output: 1.5 },
  'codestral-latest':     { input: 0.3, output: 0.9 },
  'mistral-embed':        { input: 0.1, output: 0 },
};

console.log('## Prompt Cost Estimates\n');
console.log('| File | Est. Tokens | Cost/call (small) | Cost/call (large) |');
console.log('|------|-------------|-------------------|-------------------|');

const promptDir = join(process.cwd(), 'src/prompts');
if (readdirSync(promptDir, { withFileTypes: true })) {
  for (const file of readdirSync(promptDir)) {
    const content = readFileSync(join(promptDir, file), 'utf-8');
    const tokens = Math.ceil(content.length / 4);
    const smallCost = (tokens / 1e6) * PRICING['mistral-small-latest'].input;
    const largeCost = (tokens / 1e6) * PRICING['mistral-large-latest'].input;

    console.log(`| ${file} | ~${tokens} | $${smallCost.toFixed(6)} | $${largeCost.toFixed(6)} |`);
  }
}

Step 4: Model Validation Gate

# .github/workflows/model-gate.yml
name: AI Model Gate

on:
  pull_request:
    paths: ['src/ai/**', 'src/prompts/**']

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: '20', cache: 'npm' }
      - run: npm ci

      - name: Validate model references
        run: |
          echo "## Model References" >> $GITHUB_STEP_SUMMARY
          grep -rn "mistral-\|codestral-\|pixtral-" src/ --include="*.ts" | \
            grep -oP "(mistral|codestral|pixtral)-[\w-]+" | sort -u | while read model; do
            echo "- \`$model\`" >> $GITHUB_STEP_SUMMARY
          done

      - name: Check for deprecated models
        run: |
          DEPRECATED="open-mistral-7b open-mixtral-8x7b mistral-tiny mistral-medium"
          for model in $DEPRECATED; do
            if grep -rq "$model" src/ --include="*.ts"; then
              echo "::warning::Deprecated model found: $model"
            fi
          done

Error Handling

Examples

Minimal Smoke Test

- name: Mistral smoke test
  env:
    MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
  run: |
    curl -sf -X POST https://api.mistral.ai/v1/chat/completions \
      -H "Authorization: Bearer $MISTRAL_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{"model":"mistral-small-latest","messages":[{"role":"user","content":"ping"}],"max_tokens":5}' \
      | jq '.choices[0].message.content'

Resources

• Mistral API Reference
• GitHub Actions Secrets

Output

• GitHub Actions workflow for prompt testing
• Regression test suite with deterministic assertions
• Cost estimation in PR summaries
• Model validation gate catching deprecated references