Quiz#
Hybrid Search#
Question 1: What type of query reveals a weakness in pure Vector Search?
A. Queries asking for summaries.
B. Queries requiring absolute accuracy in wording, like proper names or product codes.
C. Queries asking about complex sentiments.
D. Queries written in multiple languages.
Answer: B
Question 2: What does the BM25 algorithm specialize in finding?
A. Contextual synonyms.
B. Hypothetical vectors.
C. Precise keywords based on frequency statistics.
D. Nodes in a graph database.
Answer: C
Question 3: BM25 is considered a refined upgrade of which classical information retrieval algorithm?
A. HNSW
B. Cosine Similarity
C. TF-IDF
D. MMR.
Answer: C
Question 4: What are the two parallel search streams combined in Hybrid Search?
A. Cross-Encoder and Bi-Encoder
B. Semantic Chunking and Recursive Chunking
C. Sparse Retriever (BM25) and Dense Retriever (Vector Search)
D. Neo4j and ChromaDB.
Answer: C
Question 5: Instead of caring about numerical scores, what does the RRF algorithm rely on?
A. Rank
B. Keyword Density
C. Vector Length
D. Document Chunk Size.
Answer: A
Question 6: Why might Vector Search miss a query specifically looking for âError 503â?
A. It ignores all numbers.
B. In vector space, these numbers may not carry much specific semantic meaning, causing it to prioritize general âerrorâ synonyms.
C. It automatically rounds numbers to the nearest tenth.
D. It assumes 503 is a typographical error.
Answer: B
Question 7: How does the âTF Saturationâ mechanism in BM25 prevent keyword spamming?
A. It completely ignores keywords that appear more than once.
B. Unlike TF-IDF, after appearing a certain number of times, appearing again hardly adds more score, asymptoting to a limit.
C. It deletes the document from the index if spam is detected.
D. It divides the score by the total word count.
Answer: B
Question 8: What is the function of the IDF (Inverse Document Frequency) principle in BM25?
A. It counts the total documents in the database.
B. It ensures short paragraphs rank lower.
C. It heavily penalizes common words and greatly rewards rare words.
D. It increases the score infinitely for repeated terms.
Answer: C
Question 9: How does âLength Normalizationâ make BM25 smarter?
A. By padding short documents with empty vectors.
B. By rating a keyword occurrence in a short paragraph higher than in a long novel where information is diluted.
C. By only accepting documents of a fixed length.
D. By translating long novels into summaries.
Answer: B
Question 10: What is the primary problem encountered when trying to fuse results from Vector Search and BM25 directly?
A. They return results in different languages.
B. The scoring scales are completely different (cosine similarity vs statistical formula) and cannot be directly added.
C. Vector Search is too slow to wait for BM25.
D. BM25 cannot run in parallel.
Answer: B
Question 11: What does the âFusionâ step accomplish in the Hybrid Search Implementation Process?
A. It translates the documents.
B. It merges the two lists (from BM25 and Vector Search) into a single list.
C. It concatenates text using a Cross-Encoder.
D. It expands the query using an LLM.
Answer: B
Question 12: What is listed as a major âConâ of implementing Hybrid Search?
A. It fails if the user uses synonyms.
B. It is more complex to deploy and consumes more resources due to running 2 parallel streams.
C. It loses the ability to match exact keywords.
D. It cannot handle multi-lingual queries.
Answer: B
Question 13: In the RRF formula 1 / (k + rank_r(d)), what is the typical value chosen for the smoothing constant k?
A. 1
B. 10
C. 60
D. 100
Answer: C
Question 14: What specific role does the smoothing constant k play in the Reciprocal Rank Fusion formula?
A. It converts cosine similarity to a percentage.
B. It acts as a hard limit on the number of retrieved documents.
C. It helps reduce score disparity between very high ranks (e.g., Top 1 vs Top 2), ensuring fairness.
D. It prevents BM25 from calculating saturation.
Answer: C
Question 15: In the RRF illustrative example, why does Doc B (Rank 2 Vector, Rank 3 BM25) beat Doc A (Rank 1 Vector, Rank 10 BM25)?
A. Because BM25 scores are weighted double by default.
B. Because RRF prioritizes documents that have high consensus from both algorithms, yielding a higher combined sum (0.0320 > 0.0307).
C. Because Doc A triggered the TF saturation penalty.
D. Because Doc B was shorter, gaining a Length Normalization bonus.
Answer: B
Question 16: What is the fundamental assumption that makes Reciprocal Rank Fusion (RRF) effective?
A. That cosine similarity is always more accurate than statistical frequency.
B. That if a document appears at a high rank in both lists, it is certainly an important document.
C. That keywords in the title are more important than in the body.
D. That sparse retrievers will eventually replace dense retrievers.
Answer: B
Question 17: If a user searches for âsymptoms of meningitisâ, how does BM25 handle the word âofâ compared to âmeningitisâ?
A. It treats them equally.
B. It drops âmeningitisâ due to complexity.
C. Through IDF, it heavily penalizes âofâ as a common word and greatly rewards âmeningitisâ as a rare word.
D. It triggers length normalization based on the word âofâ.
Answer: C
Question 18: In the âGalaxyâ keyword spam example, why does TF-IDF fail compared to BM25?
A. TF-IDF cannot process the word âGalaxyâ.
B. TF-IDF scores increase linearly infinitely with repetition, giving Doc A an overwhelming win, while BM25 saturation stops the score from increasing further.
C. TF-IDF strictly relies on document length normalization.
D. BM25 uses vector similarity to realize it is spam.
Answer: B
Question 19: Which Search method is described as âPoor at matching exact keywords, hard to explain resultsâ?
A. BM25
B. Graph Traversal
C. Vector Search
D. Hybrid Search.
Answer: C
Question 20: Which Search method âDoes not understand context, fails if user uses different synonyms than the textâ?
A. Vector Search
B. Hybrid Search
C. MMR
D. BM25.
Answer: D