MidRAG & retrieval

What are hybrid search and reranking, and when does a RAG system actually need them?

What they are really testing: The standard second-generation retrieval stack. This checks you know what each stage fixes, exact-match blindness and precision-at-top respectively, rather than cargo-culting a pipeline diagram.

A real interview question

What are hybrid search and reranking, and when does a RAG system actually need them?

What most people say

drag me

Hybrid search combines keyword and semantic search for better results, and reranking reorders results with a better model, both are best practices for RAG.

Both definitions are right and the reasoning is absent. "Best practice" is the tell: each stage adds latency, cost and operational surface, and the decision to add them should come from a measured retrieval failure, not a diagram.

The follow-ups they ask next

  • Why is a cross-encoder more accurate than embedding similarity at all?

    Bi-encoders compress query and document into vectors independently, then compare. A cross-encoder attends across both texts jointly, so it can match specifics that compression lost.

  • Your reranker added 150ms and users notice. Options?

    Rerank fewer candidates, use a lighter reranker, run it only for query types that need it, or overlap it with other pipeline stages. Or revisit whether precision at 5 was really the bottleneck.

What the interviewer is listening for

  • Maps hybrid to exact-match blindness with concrete examples
  • Knows the retrieve-wide-then-rerank-narrow pattern and its latency price
  • Adds stages only when a retrieval eval shows the specific failure

What sinks the answer

  • Recommends the full stack everywhere as best practice
  • Cannot say what reranking fixes that retrieval missed
  • No measurement distinguishing recall from precision problems

If you genuinely do not know

Say this instead of freezing. Reasoning out loud from what you do know beats silence every single time, and a good interviewer is listening for exactly that.

Hybrid fixes [dense retrieval missing exact codes and names] by [fusing BM25 with vector search]. Reranking fixes [mediocre ordering in the top candidates] by [cross-encoding query and document together, retrieve 50, rerank to 5]. Add each [when a retrieval eval shows its failure], not by default.

Keep going with rag & retrieval

All 57 ai engineer questions

Knowing the answer is not the same as recalling it under pressure

Sign in to send the questions you fumble to spaced recall, so they come back right before you would forget them, and learn the concepts behind them with hands-on labs.

Start free