Retrieval strategies (similarity, MMR, hybrid) in Agentic AI - ML Experiment: Train & Evaluate

Experiment - Retrieval strategies (similarity, MMR, hybrid)

Problem:You have a document search system that returns results based on similarity scores. Currently, it returns many similar documents that are redundant. This reduces the usefulness of the search results.

Current Metrics:Precision@5: 80%, Diversity score: 0.3 (low diversity)

Issue:The retrieval strategy overemphasizes similarity, causing redundant results and low diversity.

Your Task

Improve the retrieval strategy to increase diversity of top 5 results while maintaining precision above 75%.

You can only modify the retrieval ranking method.

You cannot change the underlying document embeddings or similarity calculation.

You must keep the system efficient for real-time queries.

Hint 1

Hint 2

Hint 3

Solution

Agentic AI

import numpy as np

def mmr(doc_embeddings, query_embedding, lambda_param=0.5, top_k=5):
    # Calculate similarity between query and documents
    sim_to_query = np.dot(doc_embeddings, query_embedding) / (
        np.linalg.norm(doc_embeddings, axis=1) * np.linalg.norm(query_embedding) + 1e-10)

    selected = []
    candidates = list(range(len(doc_embeddings)))

    while len(selected) < top_k and candidates:
        mmr_scores = []
        for idx in candidates:
            if not selected:
                diversity = 0
            else:
                sim_to_selected = max(
                    np.dot(doc_embeddings[idx], doc_embeddings[s]) /
                    (np.linalg.norm(doc_embeddings[idx]) * np.linalg.norm(doc_embeddings[s]) + 1e-10)
                    for s in selected
                )
                diversity = sim_to_selected
            score = lambda_param * sim_to_query[idx] - (1 - lambda_param) * diversity
            mmr_scores.append((score, idx))

        mmr_scores.sort(reverse=True)
        best_score, best_idx = mmr_scores[0]
        selected.append(best_idx)
        candidates.remove(best_idx)

    return selected

# Example usage:
# doc_embeddings: numpy array of shape (num_docs, embedding_dim)
# query_embedding: numpy array of shape (embedding_dim,)

# Assume doc_embeddings and query_embedding are given

# selected_indices = mmr(doc_embeddings, query_embedding, lambda_param=0.7, top_k=5)

# This returns indices of documents balancing similarity and diversity.

Implemented Maximal Marginal Relevance (MMR) to select documents.

Added a lambda parameter to balance similarity to query and diversity among selected documents.

Replaced pure similarity ranking with MMR-based ranking to reduce redundancy.

Results Interpretation

Before: Precision@5 = 80%, Diversity = 0.3 (low diversity, many similar results)

After: Precision@5 = 78%, Diversity = 0.65 (higher diversity, less redundancy)

Using MMR helps balance relevance and diversity in retrieval results, reducing redundancy while keeping precision high.

Bonus Experiment

Try a hybrid retrieval strategy that first filters top 10 documents by similarity, then applies MMR to select the final 5 results.

💡 Hint

This can improve efficiency by reducing the candidate set before applying the more complex MMR ranking.

Practice

1. Which retrieval strategy focuses on ranking results purely based on how close they are to the query?

easy

A. Random retrieval

B. Maximal Marginal Relevance (MMR)

C. Similarity-based retrieval

D. Hybrid retrieval

Solution

1. Step 1: Understand similarity-based retrieval

   Similarity-based retrieval ranks results by how close or similar they are to the query, focusing only on relevance.

2. Step 2: Compare with other strategies

   MMR balances relevance and diversity, hybrid combines methods, and random is unrelated.

3. Final Answer:

   Similarity-based retrieval -> Option C

4. Quick Check:

   Similarity = closeness only [OK]

Hint: Similarity means closest match only [OK]

Common Mistakes:

• Confusing MMR with similarity
• Thinking hybrid is only similarity
• Choosing random as a valid strategy

2. Which of the following is the correct way to describe Maximal Marginal Relevance (MMR)?

easy

A. Combines all retrieval methods without weighting

B. Ranks results by random selection

C. Only uses keyword matching

D. Balances relevance and diversity in retrieval

Solution

1. Step 1: Define MMR

   MMR is designed to balance relevance to the query and diversity among the results to avoid redundancy.

2. Step 2: Eliminate incorrect options

   Random selection is unrelated, keyword matching is too narrow, and combining without weighting is not MMR.

3. Final Answer:

   Balances relevance and diversity in retrieval -> Option D

4. Quick Check:

   MMR = relevance + diversity [OK]

Hint: MMR mixes relevance with diversity [OK]

Common Mistakes:

• Thinking MMR is random
• Assuming MMR uses only keywords
• Believing MMR combines methods blindly

3. Given the following pseudo-code for a hybrid retrieval method combining similarity and MMR scores:

results = []
for doc in documents:
    sim_score = similarity(query, doc)
    mmr_score = mmr(query, doc, results)
    combined_score = 0.6 * sim_score + 0.4 * mmr_score
    results.append((doc, combined_score))
results.sort(key=lambda x: x[1], reverse=True)
print([doc for doc, score in results[:3]])

What does this code output?

medium

A. Top 3 documents ranked by combined similarity and MMR scores

B. Top 3 documents ranked by similarity score only

C. Top 3 documents ranked by MMR score only

D. Random 3 documents from the list

Solution

1. Step 1: Analyze score calculation

   The code calculates a combined score using 60% similarity and 40% MMR for each document.

2. Step 2: Understand sorting and output

   Documents are sorted by this combined score in descending order, then top 3 are printed.

3. Final Answer:

   Top 3 documents ranked by combined similarity and MMR scores -> Option A

4. Quick Check:

   Hybrid = combined scores [OK]

Hint: Check weighted sum and sorting for final ranking [OK]

Common Mistakes:

• Ignoring MMR score in combined score
• Assuming sorting by similarity only
• Thinking output is random

4. Consider this buggy code snippet for MMR retrieval:

def mmr(query, docs, selected):
    scores = []
    for doc in docs:
        relevance = similarity(query, doc)
        diversity = min([similarity(doc, s) for s in selected])
        score = relevance - 0.5 * diversity
        scores.append((doc, score))
    return max(scores, key=lambda x: x[1])[0]

What is the main error causing a crash when selected is empty?

medium

A. Using min() on an empty list causes an error

B. Incorrect use of max() function

C. Missing return statement

D. Similarity function is undefined

Solution

1. Step 1: Identify cause of crash

   When selected is empty, the list inside min() is empty, causing a ValueError.

2. Step 2: Understand min() behavior

   min() cannot operate on empty lists, so the code crashes at that line.

3. Final Answer:

   Using min() on an empty list causes an error -> Option A

4. Quick Check:

   min(empty list) = error [OK]

Hint: Check min() on empty lists for errors [OK]

Common Mistakes:

• Blaming max() instead of min()
• Ignoring empty list edge case
• Assuming similarity is undefined

5. You want to improve a search system by combining similarity and MMR retrieval. Which approach best balances relevance and diversity in the final results?

hard

A. Use MMR with a diversity weight of zero

B. Combine similarity and MMR scores with adjustable weights

C. Use only similarity scores to rank results

D. Randomly shuffle results after similarity ranking

Solution

1. Step 1: Understand the goal

   Balancing relevance and diversity requires combining both similarity and MMR scores meaningfully.

2. Step 2: Evaluate options

   Using only similarity or zero diversity weight ignores diversity; random shuffling loses relevance order.

3. Step 3: Best approach

   Combining similarity and MMR with adjustable weights allows tuning the balance effectively.

4. Final Answer:

   Combine similarity and MMR scores with adjustable weights -> Option B

5. Quick Check:

   Hybrid weighted combination = best balance [OK]

Hint: Adjust weights to balance relevance and diversity [OK]

Common Mistakes:

• Ignoring diversity by using similarity only
• Setting diversity weight to zero in MMR
• Randomizing results without scoring