Retrieval evaluation using LLM-as-a-judge via Pydantic AI

1
%pip install -U cohere pydantic-ai

1
import requests
2
import cohere
3
import pandas as pd
4
from pydantic_ai import Agent
5
from pydantic_ai.models import KnownModelName
6
from collections import Counter
7
8
import os
9
co = cohere.ClientV2(os.getenv("COHERE_API_KEY"))

1
import nest_asyncio
2
nest_asyncio.apply()

1
import requests
2
3
def search_wikipedia(query, limit=10):
4
    url = "https://en.wikipedia.org/w/api.php"
5
    params = {
6
        'action': 'query',
7
        'list': 'search',
8
        'srsearch': query,
9
        'format': 'json',
10
        'srlimit': limit
11
    }
12
13
    response = requests.get(url, params=params)
14
    data = response.json()
15
    
16
    # Format the results
17
    results = []
18
    for item in data['query']['search']:
19
        results.append({
20
            "title": item["title"],
21
            "snippet": item["snippet"].replace("<span class=\"searchmatch\">", "").replace("</span>", ""),
22
        })
23
            
24
    return results

1
# Generate 10 questions about geography to test the Wikipedia search
2
geography_questions = [
3
    "What is the capital of France?",
4
    "What is the longest river in the world?",
5
    "What is the largest desert in the world?",
6
    "What is the highest mountain peak on Earth?",
7
    "What are the major tectonic plates?",
8
    "What is the Ring of Fire?",
9
    "What is the largest ocean on Earth?",
10
    "What are the Seven Wonders of the Natural World?",
11
    "What causes the Northern Lights?",
12
    "What is the Great Barrier Reef?"
13
]

1
# Run search_wikipedia for each question
2
results = []
3
4
for question in geography_questions:
5
    question_results = search_wikipedia(question, limit=10)
6
    
7
    # Format the results as requested
8
    formatted_results = []
9
    for item in question_results:
10
        formatted_result = f"{item['title']}\n{item['snippet']}"
11
        formatted_results.append(formatted_result)
12
    
13
    # Add to the results list
14
    results.append({
15
        "question": question,
16
        "search_results": formatted_results
17
    })

1
# Rerank the search results for each question
2
top_n = 3
3
results_reranked_top_n = []
4
5
for item in results:
6
    question = item["question"]
7
    documents = item["search_results"]
8
    
9
    # Rerank the documents using Cohere
10
    reranked = co.rerank(
11
        model="rerank-v3.5",
12
        query=question,
13
        documents=documents,
14
        top_n=top_n  # Get top 3 results
15
    )
16
    
17
    # Format the reranked results
18
    top_results = []
19
    for result in reranked.results:
20
        top_results.append(documents[result.index])
21
    
22
    # Add to the reranked results list
23
    results_reranked_top_n.append({
24
        "question": question,
25
        "search_results": top_results
26
    })
27
28
# Print a sample of the reranked results
29
print(f"Original question: {results_reranked_top_n[0]['question']}")
30
print(f"Top 3 reranked results:")
31
for i, result in enumerate(results_reranked_top_n[0]['search_results']):
32
    print(f"\n{i+1}. {result}")

1
results_top_n = []
2
3
for item in results:
4
    results_top_n.append({
5
        "question": item["question"],
6
        "search_results": item["search_results"][:top_n]
7
    })
8
    
9
# Print a sample of the top_n results (without reranking)
10
print(f"Original question: {results_top_n[0]['question']}")
11
print(f"Top {top_n} results (without reranking):")
12
for i, result in enumerate(results_top_n[0]['search_results']):
13
    print(f"\n{i+1}. {result}")

1
# System prompt for the AI evaluator
2
SYSTEM_PROMPT = """
3
You are an AI search evaluator. You will compare search results from two engines and
4
determine which set provides more relevant and diverse information. You will only
5
answer with the verdict rather than explaining your reasoning; simply say "Engine A" or
6
"Engine B".
7
"""
8
9
# Prompt template for each evaluation
10
PROMPT_TEMPLATE = """
11
For the following question, which search engine provides more relevant results?
12
13
## Question:
14
{query}
15
16
## Engine A:
17
{engine_a_results}
18
19
## Engine B:
20
{engine_b_results}
21
"""
22
23
def format_results(results):
24
    """Format search results in a readable way"""
25
    formatted = []
26
    for i, result in enumerate(results):
27
        formatted.append(f"Result {i+1}: {result[:200]}...")
28
    return "\n\n".join(formatted)
29
30
def judge_query(query, engine_a_results, engine_b_results, model_name):
31
    """Use a single model to judge which engine has better results"""
32
    agent = Agent(model_name, system_prompt=SYSTEM_PROMPT)
33
    
34
    # Format the results
35
    engine_a_formatted = format_results(engine_a_results)
36
    engine_b_formatted = format_results(engine_b_results)
37
    
38
    # Create the prompt
39
    prompt = PROMPT_TEMPLATE.format(
40
        query=query,
41
        engine_a_results=engine_a_formatted,
42
        engine_b_results=engine_b_formatted
43
    )
44
    
45
    # Get the model's judgment
46
    response = agent.run_sync(prompt)
47
    return response.data
48
49
def evaluate_search_results(reranked_results, regular_results, models):
50
    """
51
    Evaluate both sets of search results using multiple models.
52
    
53
    Args:
54
        reranked_results: List of dictionaries with 'question' and 'search_results'
55
        regular_results: List of dictionaries with 'question' and 'search_results'
56
        models: List of model names to use as judges
57
    
58
    Returns:
59
        DataFrame with evaluation results
60
    """
61
    # Prepare data structure for results
62
    evaluation_results = []
63
    
64
    # Evaluate each query
65
    for i in range(len(reranked_results)):
66
        query = reranked_results[i]['question']
67
        engine_a_results = reranked_results[i]['search_results']  # Reranked results
68
        engine_b_results = regular_results[i]['search_results']   # Regular results
69
        
70
        # Get judgments from each model
71
        judgments = []
72
        for model in models:
73
            judgment = judge_query(query, engine_a_results, engine_b_results, model)
74
            judgments.append(judgment)
75
        
76
        # Determine winner by majority vote
77
        votes = Counter(judgments)
78
        if votes["Engine A"] > votes["Engine B"]:
79
            winner = "Engine A"
80
        elif votes["Engine B"] > votes["Engine A"]:
81
            winner = "Engine B"
82
        else:
83
            winner = "Tie"
84
        
85
        # Add results for this query
86
        row = [query] + judgments + [winner]
87
        evaluation_results.append(row)
88
    
89
    # Create DataFrame
90
    column_names = ["question"] + [f"judge_{i+1} ({model})" for i, model in enumerate(models)] + ["winner"]
91
    df = pd.DataFrame(evaluation_results, columns=column_names)
92
    
93
    return df

1
# Define the search engines
2
engine_a = results_reranked_top_n
3
engine_b = results_top_n
4
5
# Define the models to use as judges
6
models = [
7
    "cohere:command-a-03-2025",
8
    "cohere:command-r-plus-08-2024",
9
    "cohere:command-r-08-2024",
10
    "cohere:c4ai-aya-expanse-32b",
11
]
12
13
# Get evaluation results
14
results_df = evaluate_search_results(engine_a, engine_b, models)
15
16
# Calculate overall statistics
17
winner_counts = Counter(results_df["winner"])
18
total_queries = len(results_df)
19
20
# Display summary of results
21
print("\nPercentage of questions won by each engine:")
22
for engine, count in winner_counts.items():
23
    percentage = (count / total_queries) * 100
24
    print(f"{engine}: {percentage:.2f}% ({count}/{total_queries})")
25
    
26
# Display dataframe
27
results_df.head()
28
29
# Save to CSV
30
results_csv = results_df.to_csv("search_results_evaluation.csv", index=False)