How do you choose between vector-only and hybrid retrieval?

Instruction: Describe when embeddings alone are enough and when lexical signals should stay in the stack.

Context: Checks whether the candidate can explain the core concept clearly and connect it to real production decisions. Describe when embeddings alone are enough and when lexical signals should stay in the stack.

Updated

Example Answer

I would compare retrieval methods on the query types the product actually receives. Dense retrieval can help with paraphrases, while lexical retrieval is useful for exact identifiers such as error E104 or a quoted policy title.

For a corpus containing both, I would test a hybrid candidate set and fuse the ranked lists with a method such as reciprocal rank fusion. I would not simply average raw BM25 and vector scores, because their scales and meanings differ. Reranking can then select the small evidence set that enters the prompt.

I would measure supporting-passage coverage and latency on both exact-match and semantic slices, then keep the simplest method that meets the bar. Hybrid adds another query and tuning work; it is worth that cost only if it fixes meaningful misses rather than improves one blended headline score.

Why this works

The answer gives a concrete query and explains why fusion is a separate design choice from running two retrievers.

Follow-up to prepare for

How would you isolate the benefit? Compare lexical-only, vector-only, and fused candidates on the same queries before changing generation.

Related Questions