Hybrid Search Quiz

Answers and explanations appear as you take the quiz.

1. What two retrieval approaches does hybrid search typically combine?
   • Sparse keyword search (e.g., BM25) and dense vector search
   • Two different large language models voting on answers
   • SQL queries and full-text indexes only
   • Image search and audio search
2. Why is Reciprocal Rank Fusion (RRF) commonly used to merge results?
   • It trains a new neural network for each query
   • It combines documents by rank position instead of raw scores, avoiding scale mismatch
   • It deletes any document that appears in both lists
   • It only keeps the single top result from each retriever
3. Which scenario most favors the sparse (keyword) component of hybrid search?
   • Finding documents that paraphrase the query with different words
   • Understanding the emotional tone of a sentence
   • Matching an exact part number or error code that must appear literally
   • Translating a query into another language
4. What is the main weakness of using dense vector search alone?
   • It cannot understand synonyms or paraphrases
   • It only works on images
   • It requires no embeddings at all
   • It can miss exact rare tokens like specific codes or surnames
5. What does a learned sparse model like SPLADE do?
   • It produces interpretable term weights using a neural network, blending sparse and dense ideas
   • It removes all keyword matching from search
   • It only stores images as vectors
   • It replaces the need for any retrieval
6. What can combining lexical and dense retrieval provide to a RAG system?
   • A guarantee that every generated statement is true
   • Potentially more varied candidate evidence, subject to evaluation of retrieval and answer quality
   • A replacement for grounding sources
   • An automatic legal review of generated answers
7. What tradeoff did BEIR report for reranking and late-interaction methods on average?
   • They always ran faster and used less compute than BM25
   • They made retrieval independent of the corpus
   • They often achieved stronger zero-shot performance but at higher computational cost
   • They guaranteed a correct generated answer
8. What did the BEIR benchmark conclude about BM25 in zero-shot retrieval?
   • It always beats every neural retriever
   • It cannot retrieve exact terms
   • It is a dense-vector model
   • It remains a robust baseline across diverse datasets