Multi-query retrieval in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What is the main advantage of using multi-query retrieval in search systems?

Imagine you want to find information on a complex topic that has several aspects. Why would a multi-query retrieval system be better than a single-query system?

A It simplifies the search by ignoring less important parts of the topic.

B It reduces the number of queries needed by combining all keywords into one long query.

C It guarantees faster retrieval by only using one query at a time.

D It allows searching multiple aspects simultaneously, improving recall and relevance.

💡 Hint

Think about how covering different angles of a topic can help find more useful results.

Multi-query retrieval sends several focused queries to cover different aspects of a topic, which helps find more relevant documents and improves recall compared to a single broad query.

❓ Predict Output

intermediate

Output of multi-query retrieval scoring aggregation

Given the following Python code that simulates scores from multiple queries for documents, what is the final aggregated score for document 'doc2'?

Prompt Engineering / GenAI

query_scores = {
    'query1': {'doc1': 0.8, 'doc2': 0.5, 'doc3': 0.3},
    'query2': {'doc1': 0.4, 'doc2': 0.7, 'doc3': 0.6},
    'query3': {'doc1': 0.3, 'doc2': 0.2, 'doc3': 0.9}
}

# Aggregate scores by summing scores from all queries per document
aggregated_scores = {}
for q_scores in query_scores.values():
    for doc, score in q_scores.items():
        aggregated_scores[doc] = aggregated_scores.get(doc, 0) + score

final_score_doc2 = aggregated_scores['doc2']
print(final_score_doc2)

A 0.9

B 1.4

C 1.2

D 1.0

💡 Hint

Sum the scores for 'doc2' from each query.

Scores for 'doc2' are 0.5 (query1) + 0.7 (query2) + 0.2 (query3) = 1.4.

❓ Model Choice

advanced

Best model architecture for multi-query retrieval embeddings

You want to build a system that generates embeddings for multiple queries to retrieve documents effectively. Which model architecture is best suited for encoding multiple queries independently but efficiently?

A A recurrent neural network that concatenates all queries into one sequence before encoding.

B A convolutional neural network that treats all queries as one long image input.

C A single shared transformer encoder that processes each query separately to produce embeddings.

D A decision tree model that classifies queries into categories before retrieval.

💡 Hint

Think about models that can handle sequences independently and produce vector representations.

A shared transformer encoder can efficiently encode each query independently into embeddings, which is ideal for multi-query retrieval systems.

❓ Hyperparameter

advanced

Choosing the best aggregation method hyperparameter for multi-query retrieval

In multi-query retrieval, you combine scores from multiple queries to rank documents. Which hyperparameter choice affects how you weight each query's contribution?

A The query weight vector that assigns importance to each query's score before aggregation.

B The dropout rate applied in the neural network layers.

C The batch size used during model training.

D The learning rate used to train the embedding model.

💡 Hint

Focus on parameters that control score combination rather than model training.

The query weight vector controls how much each query's score influences the final aggregated score, directly affecting retrieval ranking.

❓ Metrics

expert

Evaluating multi-query retrieval with combined metrics

You have retrieval results from multiple queries combined to rank documents. Which evaluation metric best captures both the relevance and ranking quality across all queries?

A Mean Reciprocal Rank (MRR) averaged over all queries.

B Accuracy of classifying documents as relevant or not.

C Mean Squared Error (MSE) between predicted and true scores.

D Confusion matrix counts aggregated over all queries.

💡 Hint

Think about metrics designed for ranking quality in retrieval tasks.

MRR measures the average rank position of the first relevant document per query, making it suitable for multi-query retrieval evaluation.