Skip to content
Quantoros

What two-stage idea underlies reranking plus hybrid retrieval?

LLM & AILevel: IntermediateRAG (Retrieval-Augmented Generation)

What two-stage idea underlies reranking plus hybrid retrieval?

Hints
  • Break the query into candidate retrieval, context selection and generation.
  • How would you locate the stage that lost a fact without using only one end-to-end metric?

Answer

Hybrid retrieval combines complementary retrieval signals, commonly lexical matching and dense vector similarity. Retrieve candidates from both channels and combine their rankings, for example with reciprocal rank fusion or an appropriately calibrated score combination. This helps when exact names or identifiers matter alongside semantic paraphrases.

Reranking is a separate second stage: a more expensive model evaluates the query jointly with each shortlisted candidate and reorders the list. The first stage targets coverage at manageable cost; the reranker targets ordering quality. A reranker cannot recover a relevant document that never entered the candidate set. Tune candidate counts and fusion against retrieval recall, ranking metrics and end-to-end latency.

Source and licence: Skillumen — Awesome AI Engineer Interview Questions · CC0-1.0 · answer: Quantoros editorial

Related questions