Best Practices

Expert tips for effective model testing, evaluation, and sustainable deployment.

---

Testing Strategy

1. Start Simple

✅ Begin with defaults:

• Model: GPT-4o Mini (A+ sustainability rating)
• Temperature: 0.7
• Max Tokens: 500

Why: Establish a baseline with a well-balanced, sustainable model.

2. Test Multiple Models

✅ Compare across providers:

Test 1: GPT-4o Mini (Azure) - Default
Test 2: Claude 3 Haiku (AWS) - Speed focus
Test 3: Gemini 2.5 Flash (Google) - Alternative
Test 4: GPT-4.1 (Azure) - Quality focus

Compare: Quality, speed, cost, and CO₂ emissions.

3. Use Real Data

✅ Test with actual use case prompts:

• Real customer questions
• Actual data to analyze
• Production-like scenarios

✅ Leverage prompt history:

• Review what prompts worked before
• Re-run successful tests
• Track performance over time

❌ Don't: Only test with toy examples.

---

Model Selection Framework

Decision Matrix

Quality Requirements:
  Critical (99%) → GPT-4.1 or GPT-5 (if absolutely needed)
  High (95%) → GPT-4o or Claude Sonnet 4
  Medium (90%) → GPT-4o Mini or GPT-4.1 Mini
  Basic (85%) → Claude 3 Haiku or Amazon Nova Lite

Sustainability Priority:
  Most sustainable → Amazon Nova Lite (A+, 0.10g CO₂)
  Very sustainable → Gemini 2.0 Flash Lite (A+, 0.12g CO₂)
  Good balance → GPT-4o Mini (A+, 0.64g CO₂)
  High quality + sustainable → GPT-4.1 (A+, 0.56g CO₂)

Hybrid Approach (Recommended)

Best Practice: Use multiple models based on task complexity

function selectModel(request) {
  const complexity = analyzeComplexity(request);
  
  switch(complexity) {
    case 'complex':
      return 'GPT-4.1';           // 10% of requests, A+ rating
    case 'standard':
      return 'GPT-4o-Mini';       // 60% of requests, A+ rating
    case 'simple':
      return 'Claude-3-Haiku';    // 30% of requests, A+ rating
  }
}

Result: All A+ sustainability ratings while meeting quality needs.

---

Evaluation Criteria

Define Success Metrics

Before testing, define:

1. Quality Threshold

   Customer Support: 90% helpful responses
   Code Generation: 95% syntactically correct
   Data Classification: 98% accuracy

2. Performance Requirements

   Response Time: < 2 seconds
   Token Usage: < 500 tokens average
   Cost per Request: < $0.01

3. Sustainability Goals

   CO₂ per Request: < 1.0g
   Sustainability Rating: A+ or A preferred
   Monthly CO₂ Budget: < 10 kg

Measure Objectively

Quantitative Metrics:

• Accuracy (for classification)
• Response time (milliseconds)
• Token count (cost proxy)
• CO₂ emissions (grams)
• Water consumption (liters)

Qualitative Assessment:

• Relevance (1-5 scale)
• Clarity (1-5 scale)
• Accuracy (1-5 scale)

---

Sample Size Requirements

Minimum Test Runs

Use Case	Min Tests	Why
Classification	50	Need statistical significance
Content Generation	10	Check consistency
Customer Support	20	Vary user intents
Data Analysis	15	Test edge cases

Testing Pattern

// ✅ Good: Multiple runs with metrics
const results = [];
for (let i = 0; i < 20; i++) {
  const result = await testModel(prompt, settings);
  results.push({
    quality: result.quality,
    responseTime: result.duration,
    tokens: result.tokens,
    co2: result.co2Grams,
    cost: result.costUsd
  });
}
analyzeResults(results);

// ❌ Poor: Single test
const result = await testModel(prompt, settings);
deployToProduction(); // Risky!

---

Sustainability Best Practices

1. Prefer A+ Rated Models

A+ rated models (lowest CO₂):

• Amazon Nova Lite (0.10g/1k)
• Gemini 2.0 Flash Lite (0.12g/1k)
• GPT-4.1 (0.56g/1k)
• GPT-4o Mini (0.64g/1k)
• Claude 3 Haiku (0.64g/1k)

2. Monitor Environmental Impact

const monthlyMetrics = {
  totalRequests: 10000,
  totalCO2Grams: 6400,    // 6.4 kg
  totalWaterLiters: 4600,
  avgCO2PerRequest: 0.64,
  sustainabilityRating: 'A+'
};

// Set sustainability alerts
if (monthlyMetrics.totalCO2Grams > 10000) { // 10 kg threshold
  alert("Monthly CO₂ budget exceeded");
}

3. Right-Size Your Models

Before: GPT-5 for all tasks

• CO₂: 13.78g per 1k tokens
• Cost: $0.020 per 1k tokens

After: Model routing

• Simple: Claude 3 Haiku (0.64g, $0.0008)
• Standard: GPT-4o Mini (0.64g, $0.0015)
• Complex: GPT-4.1 (0.56g, $0.012)

Savings: 95% CO₂ reduction, 90% cost reduction

---

Cost Management

Monitor Token Usage

Track these metrics:

const metrics = {
  avgInputTokens: 150,
  avgOutputTokens: 300,
  totalTokens: 450,
  costPerRequest: 0.008,
  co2PerRequest: 0.29,     // grams
  requestsPerDay: 1000,
  dailyCost: 8.00,
  dailyCO2: 290             // grams
};

Optimization Strategies

1. Prompt Optimization

   ❌ Verbose: "I would like you to please help me by..."
   ✅ Concise: "Summarize this text:"
   
   Savings: 20-30% fewer input tokens

2. Token Limits

   ❌ No limit: Average 800 tokens
   ✅ Set limit: 400 tokens
   
   Savings: 50% on output costs and CO₂

3. Model Routing

   ❌ All GPT-5: $200/month, 137 kg CO₂
   ✅ Mixed models: $30/month, 7 kg CO₂
   
   Savings: 85% cost, 95% CO₂

---

Quality Assurance

Pre-Deployment Checklist

• Tested with 20+ real prompts
• Success rate > threshold
• Response time acceptable
• Cost within budget
• CO₂ within sustainability goals
• Edge cases handled
• Error handling tested
• Monitoring in place

Post-Deployment Monitoring

// Monitor production metrics
setInterval(() => {
  const metrics = collectMetrics();
  
  if (metrics.quality < qualityThreshold) {
    alert("Quality degradation detected");
  }
  
  if (metrics.dailyCost > costBudget) {
    alert("Budget exceeded");
  }
  
  if (metrics.dailyCO2 > co2Budget) {
    alert("CO₂ budget exceeded - consider model optimization");
  }
}, 3600000); // Check hourly

---

Common Mistakes

❌ Mistake 1: Using Premium Models for Simple Tasks

Problem: GPT-5 for classification = overkill

• Cost: $0.020 per 1k tokens
• CO₂: 13.78g per 1k tokens

Solution: Use Claude 3 Haiku or Amazon Nova Lite

• Cost: $0.0005-0.0008 per 1k tokens
• CO₂: 0.10-0.64g per 1k tokens
• Savings: 97% cost, 95% CO₂

❌ Mistake 2: Ignoring Sustainability Metrics

Problem: Focus only on cost, miss environmental impact

Solution: Track CO₂ alongside cost

• Use A+ rated models when possible
• Set monthly CO₂ budgets
• Report on sustainability metrics

❌ Mistake 3: Single Model Strategy

Problem: Using one model for everything

Solution: Implement model routing

// Route based on complexity and sustainability
function selectModel(task) {
  if (task.complexity === 'high' && task.qualityCritical) {
    return 'GPT-4.1';      // A+ rating, high quality
  }
  return 'GPT-4o-Mini';    // A+ rating, balanced
}

❌ Mistake 4: Skipping Documentation

Problem: Can't reproduce or explain results

Solution: Document everything

• Test configurations
• Success criteria
• Sustainability metrics
• Rationale for model choices

Pro Tip: GenAI Explorer's prompt history automatically documents your tests!

---

Testing Template

Use This Structure

## Test: [Test Name]

**Date:** [Date]
**Tester:** [Name]

### Configuration
- Model: GPT-4o Mini
- Temperature: 0.7
- Max Tokens: 500

### Test Results
- Success Rate: 92%
- Avg Response Time: 1.2s
- Avg Cost: $0.0007
- Avg CO₂: 0.29g
- Sustainability Rating: A+

### Comparison with Alternative
| Metric | GPT-4o Mini | Claude 3 Haiku |
|--------|-------------|----------------|
| Quality | 92% | 88% |
| Speed | 1.2s | 0.8s |
| Cost | $0.0007 | $0.0004 |
| CO₂ | 0.29g | 0.29g |

### Decision
[x] Deploy GPT-4o Mini (better quality)
[ ] Try different model

---

Advanced Tips

1. A/B Testing with Sustainability

Compare approaches including environmental impact:

Group A: GPT-4o Mini, temp 0.5
Group B: Claude 3 Haiku, temp 0.7

Compare: Quality, Cost, Speed, CO₂ emissions
Winner: Based on your priorities

2. Sustainability-First Selection

When multiple models meet quality requirements:

function selectMostSustainableModel(qualifiedModels) {
  return qualifiedModels.sort((a, b) => 
    a.co2PerToken - b.co2PerToken
  )[0]; // Return lowest CO₂ option
}

3. Prompt Versioning with Metrics

Track improvements across versions:

v1.0: Original prompt → 500 tokens, $0.0008, 0.32g CO₂
v1.1: Optimized prompt → 350 tokens, $0.0005, 0.22g CO₂
v1.2: Added examples → 400 tokens, $0.0006, 0.26g CO₂

Best: v1.1 (30% CO₂ reduction)

---

Related Documentation

• Model Comparison - Full model reference with sustainability
• Sustainability Guide - Deep dive into environmental impact
• Cost Optimization - Reduce spending
• Parameters Reference - All settings
• Sample Prompts - Templates

---

Follow these best practices for successful, sustainable AI model testing and deployment.