AI-300 Practice Question 29: Hybrid Search Exact Match
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 3 · question 9 of 10
A retrieval-augmented assistant over an industrial parts catalog handles conceptual questions well, but when a technician pastes an exact part number such as RX-4410-B, the retriever returns similar-sounding parts and misses the exact record. The team is not willing to give up the conceptual matching that works today. What should you change in the query?
Show answer and explanation
Answer:
"Hybrid search combines the strengths of vector search and keyword search", and Microsoft names this exact failure mode: "Some scenarios, such as querying over product codes, highly specialized jargon, dates, and people's names, perform better with keyword search because it can identify exact matches." A hybrid query "is a single query request that includes both search and vectorQueries query parameters" and "merges results from each query by using Reciprocal Rank Fusion (RRF)", so the lexical channel surfaces RX-4410-B while the vector channel keeps answering the conceptual questions (option D).
- A is wrong because a larger k returns more approximate neighbors from the same embedding space. Without a lexical channel there is still nothing matching the literal token, so the exact record may simply not be nearby in vector space at any k. - B is wrong because a stronger embedding model improves semantic similarity, which is already working. Embeddings are the documented weak tool for exact identifier matching, and re-embedding the entire catalog is an expensive way to not fix the problem. - C is wrong because a lower threshold admits *weaker* matches. That adds more near-miss parts to the results (the very noise the technician is complaining about) without promoting the exact record.
Reference: Hybrid search overview - Azure AI Search