1/**
2 * Pairwise Comparison Example
3 * 
4 * Demonstrates how to compare two responses and pick the better one.
5 * 
6 * Run: npx tsx examples/pairwise-comparison.ts
7 */
8 
9import 'dotenv/config';
10import { EvaluatorAgent } from '../src/agents/evaluator.js';
11import { validateConfig } from '../src/config/index.js';
12 
13async function main() {
14  validateConfig();
15 
16  const agent = new EvaluatorAgent();
17 
18  console.log('=== Pairwise Comparison Example ===\n');
19 
20  const prompt = 'Explain the benefits of regular exercise';
21 
22  const responseA = `
23    Regular exercise offers numerous health benefits that affect both body and mind.
24    
25    Physical benefits include:
26    - Improved cardiovascular health and reduced heart disease risk
27    - Stronger muscles and bones
28    - Better weight management
29    - Enhanced immune function
30    
31    Mental benefits include:
32    - Reduced stress and anxiety
33    - Improved mood through endorphin release
34    - Better sleep quality
35    - Enhanced cognitive function
36    
37    The CDC recommends at least 150 minutes of moderate aerobic activity per week,
38    plus muscle-strengthening activities twice weekly.
39  `;
40 
41  const responseB = `
42    Working out is really good for you. It makes you healthier and feel better.
43    You should try to exercise regularly if you can. Many people find that
44    going to the gym or running helps them stay in shape.
45  `;
46 
47  console.log('Prompt:', prompt);
48  console.log('\n--- Response A ---');
49  console.log(responseA.trim());
50  console.log('\n--- Response B ---');
51  console.log(responseB.trim());
52  console.log('\n--- Comparison Results ---\n');
53 
54  const result = await agent.compare({
55    responseA,
56    responseB,
57    prompt,
58    criteria: ['accuracy', 'completeness', 'actionability', 'clarity'],
59    allowTie: true,
60    swapPositions: true // Mitigate position bias
61  });
62 
63  if (result.success) {
64    console.log(`Winner: Response ${result.winner}`);
65    console.log(`Confidence: ${(result.confidence * 100).toFixed(0)}%`);
66    
67    if (result.positionConsistency) {
68      console.log(`Position Consistency: ${result.positionConsistency.consistent ? 'Yes' : 'No'}`);
69    }
70 
71    console.log('\nPer-Criterion Results:');
72    result.comparison.forEach(c => {
73      console.log(`\n  ${c.criterion}:`);
74      console.log(`    Winner: ${c.winner}`);
75      console.log(`    A: ${c.aAssessment}`);
76      console.log(`    B: ${c.bAssessment}`);
77    });
78 
79    console.log('\nKey Differentiators:');
80    result.differentiators.forEach(d => console.log(`  - ${d}`));
81 
82    console.log('\nResponse A Analysis:');
83    console.log('  Strengths:', result.analysis.responseA.strengths.join(', '));
84    console.log('  Weaknesses:', result.analysis.responseA.weaknesses.join(', '));
85 
86    console.log('\nResponse B Analysis:');
87    console.log('  Strengths:', result.analysis.responseB.strengths.join(', '));
88    console.log('  Weaknesses:', result.analysis.responseB.weaknesses.join(', '));
89 
90    console.log(`\nEvaluation Time: ${result.metadata.evaluationTimeMs}ms`);
91  } else {
92    console.error('Comparison failed');
93  }
94}
95 
96main().catch(console.error);
97 
98