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

// Test with cache disabled to get fresh AI responses
import { default as IAEngine } from './src/DRS/services/IAEngine.js';

async function testNoCacheStrictScoring() {
    console.log('🎯 FRESH TEST - No cache, strict scoring validation\n');

    // Create new engine instance to avoid cache
    const engine = new IAEngine({
        defaultProvider: 'openai',
        fallbackProviders: ['deepseek']
    });

    // Clear any existing cache
    engine.cache?.clear();

    await new Promise(resolve => setTimeout(resolve, 1000));

    console.log('🧪 Testing wrong answers (should get <20 points):\n');

    // Test 1: Text comprehension - completely wrong
    try {
        console.log('1️⃣ Text Analysis: "Elephants are purple" for Amazon rainforest');
        const result1 = await engine.validateComprehension(
            'The Amazon rainforest is the largest tropical rainforest in the world.',
            'Elephants are purple animals',
            { exerciseType: 'text' }
        );
        console.log(`   Score: ${result1.score} (should be <20: ${result1.score < 20 ? 'PASS' : 'FAIL'})`);
        console.log(`   Provider: ${result1.provider}`);
    } catch (error) {
        console.log(`   ❌ Failed: ${error.message}`);
    }

    await new Promise(resolve => setTimeout(resolve, 3000));

    // Test 2: Translation - completely wrong
    try {
        console.log('\n2️⃣ Translation: "Pizza spaghetti" for "Good morning"');
        const result2 = await engine.validateTranslation(
            'Good morning',
            'Pizza spaghetti',
            { fromLang: 'en', toLang: 'fr' }
        );
        console.log(`   Score: ${result2.score} (should be <20: ${result2.score < 20 ? 'PASS' : 'FAIL'})`);
        console.log(`   Provider: ${result2.provider}`);
    } catch (error) {
        console.log(`   ❌ Failed: ${error.message}`);
    }

    await new Promise(resolve => setTimeout(resolve, 3000));

    console.log('\n🧪 Testing correct answers (should get >70 points):\n');

    // Test 3: Good comprehension
    try {
        console.log('3️⃣ Text Analysis: "Amazon is the biggest rainforest" for Amazon rainforest');
        const result3 = await engine.validateComprehension(
            'The Amazon rainforest is the largest tropical rainforest in the world.',
            'Amazon is the biggest rainforest',
            { exerciseType: 'text' }
        );
        console.log(`   Score: ${result3.score} (should be >70: ${result3.score > 70 ? 'PASS' : 'FAIL'})`);
        console.log(`   Provider: ${result3.provider}`);
    } catch (error) {
        console.log(`   ❌ Failed: ${error.message}`);
    }

    await new Promise(resolve => setTimeout(resolve, 3000));

    // Test 4: Good translation
    try {
        console.log('\n4️⃣ Translation: "Bonjour" for "Good morning"');
        const result4 = await engine.validateTranslation(
            'Good morning',
            'Bonjour',
            { fromLang: 'en', toLang: 'fr' }
        );
        console.log(`   Score: ${result4.score} (should be >70: ${result4.score > 70 ? 'PASS' : 'FAIL'})`);
        console.log(`   Provider: ${result4.provider}`);
    } catch (error) {
        console.log(`   ❌ Failed: ${error.message}`);
    }

    console.log('\n🎯 FRESH TEST SUMMARY:');
    console.log('✅ If all tests PASS, the strict scoring is working correctly');
    console.log('❌ If tests FAIL, the prompts need further adjustment');
}

testNoCacheStrictScoring().catch(console.error);