Class_generator/tests/ai-validation/test-no-cache-consistency.js

// Test de cohérence SANS cache - Prompts complètement différents
import { default as IAEngine } from './src/DRS/services/IAEngine.js';

async function testNoCacheConsistency() {
    console.log('🔄 TEST DE COHÉRENCE SANS CACHE - Prompts uniques\n');
    console.log('================================================\n');

    // Test cas spécifiques avec délais pour éviter cache
    const testRounds = [
        {
            name: 'Round 1 - Mauvaise réponse',
            content: 'The Eiffel Tower is located in Paris, France.',
            answer: 'Cats are purple and live on the moon',
            expectedRange: [0, 30]
        },
        {
            name: 'Round 2 - Bonne réponse',
            content: 'Tokyo is the capital city of Japan with over 13 million people.',
            answer: 'Tokyo is Japan\'s capital and has more than 13 million inhabitants',
            expectedRange: [70, 100]
        },
        {
            name: 'Round 3 - Mauvaise traduction',
            original: 'I love books',
            translation: 'Monkey banana computer',
            expectedRange: [0, 30]
        },
        {
            name: 'Round 4 - Bonne traduction',
            original: 'The weather is nice today',
            translation: 'Le temps est beau aujourd\'hui',
            expectedRange: [70, 100]
        },
        {
            name: 'Round 5 - Autre mauvaise réponse',
            content: 'New York City has five boroughs: Manhattan, Brooklyn, Queens, Bronx, and Staten Island.',
            answer: 'Fish swim in chocolate rivers',
            expectedRange: [0, 30]
        }
    ];

    const results = [];

    for (let i = 0; i < testRounds.length; i++) {
        const round = testRounds[i];
        console.log(`🧪 ${round.name}`);

        // Créer une nouvelle instance pour éviter le cache
        const engine = new IAEngine({
            defaultProvider: 'openai',
            fallbackProviders: ['deepseek']
        });

        try {
            let result;
            if (round.content) {
                // Test de compréhension
                console.log(`   Texte: "${round.content}"`);
                console.log(`   Réponse: "${round.answer}"`);
                result = await engine.validateComprehension(round.content, round.answer, {
                    exerciseType: `comprehension-test-${Date.now()}-${i}`
                });
            } else {
                // Test de traduction
                console.log(`   Original: "${round.original}"`);
                console.log(`   Traduction: "${round.translation}"`);
                result = await engine.validateTranslation(round.original, round.translation, {
                    fromLang: 'en',
                    toLang: 'fr',
                    testId: `translation-test-${Date.now()}-${i}`
                });
            }

            const [min, max] = round.expectedRange;
            const inRange = result.score >= min && result.score <= max;

            console.log(`   📊 Score: ${result.score} (attendu: ${min}-${max})`);
            console.log(`   ✅ Dans la plage: ${inRange ? 'OUI' : 'NON'}`);
            console.log(`   🤖 Provider: ${result.provider}`);
            console.log(`   💬 Feedback: ${result.feedback?.substring(0, 80)}...\n`);

            results.push({
                name: round.name,
                score: result.score,
                expected: round.expectedRange,
                inRange: inRange,
                provider: result.provider
            });

        } catch (error) {
            console.log(`   ❌ Erreur: ${error.message}\n`);
            results.push({
                name: round.name,
                score: 'ERROR',
                expected: round.expectedRange,
                inRange: false,
                error: error.message
            });
        }

        // Délai important pour éviter rate limiting et cache
        await new Promise(resolve => setTimeout(resolve, 5000));
    }

    // Analyse finale
    console.log('📊 ANALYSE FINALE SANS CACHE:');
    console.log('============================\n');

    const validResults = results.filter(r => typeof r.score === 'number');
    const badAnswerResults = validResults.filter(r => r.expected[1] <= 30); // Mauvaises réponses
    const goodAnswerResults = validResults.filter(r => r.expected[0] >= 70); // Bonnes réponses

    console.log('🔴 Mauvaises réponses (devraient avoir <30 points):');
    badAnswerResults.forEach(r => {
        console.log(`   ${r.inRange ? '✅' : '❌'} ${r.name}: ${r.score} points`);
    });

    console.log('\n🟢 Bonnes réponses (devraient avoir >70 points):');
    goodAnswerResults.forEach(r => {
        console.log(`   ${r.inRange ? '✅' : '❌'} ${r.name}: ${r.score} points`);
    });

    const passedTests = validResults.filter(r => r.inRange).length;
    const totalTests = validResults.length;

    console.log(`\n🎯 RÉSULTAT FINAL:`);
    console.log(`   Tests réussis: ${passedTests}/${totalTests} (${Math.round((passedTests/totalTests)*100)}%)`);

    if (passedTests === totalTests) {
        console.log('\n🎉 SYSTÈME PARFAIT!');
        console.log('✅ Toutes les mauvaises réponses reçoivent des scores bas');
        console.log('✅ Toutes les bonnes réponses reçoivent des scores élevés');
        console.log('✅ Le scoring IA fonctionne correctement');
    } else if (passedTests >= totalTests * 0.8) {
        console.log('\n✅ SYSTÈME ACCEPTABLE');
        console.log('La plupart des tests passent, système utilisable');
    } else {
        console.log('\n❌ SYSTÈME PROBLÉMATIQUE');
        console.log('Trop de scores inappropriés, besoin d\'ajustements');
    }

    return results;
}

testNoCacheConsistency().catch(console.error);