AI-103 Practice Question 10: Hybrid Search Semantic Ranker
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 1 · question 10 of 10
A RAG assistant grounded on 60,000 chunks of internal policy prose is retrieving the wrong passages in two distinct ways. Queries that quote an exact clause identifier such as "FIN-4471" return chunks that are merely about the same subject, because pure vector search matches on meaning rather than the literal token. Queries phrased as questions retrieve topically correct chunks, but the single most on-point passage often lands eighth or ninth instead of first. The index already carries both text and vector fields.
Which change addresses both failures?
Show answer and explanation
Answer:
The two failures need two different fixes, and this option supplies both. Hybrid search runs the keyword and vector queries together and fuses them with Reciprocal Rank Fusion (RRF), so a literal token like FIN-4471 is matched by the BM25 side that vector similarity alone misses. The semantic ranker then adds L2 ranking: a secondary pass over the RRF-ranked result set using multilingual deep learning models adapted from Microsoft Bing, which reorders by semantic meaning and promotes the genuinely most relevant passage to the top. Information-rich prose such as policy documentation is the content type the ranker benefits most.
- B is wrong because it treats a ranking problem as a volume problem. Flooding the prompt with 50 chunks raises token cost and latency and dilutes the grounding context; it does nothing for the exact-identifier miss, since the right chunk may never be retrieved at all. - C is wrong because it trades one failure for the other. Keyword-only search fixes exact identifiers and loses the semantic matching that makes natural-language questions work. - D is wrong because a different embedding model still ranks by vector similarity. Raising the similarity threshold makes the exact-identifier case worse by discarding borderline matches, and the reranking problem is untouched.
Reference: Semantic ranking in Azure AI Search