techniques
Reranking
A second-stage scoring model that reorders the top 50 retrieval candidates before sending the top 5 to the synthesis model.
Definition
Reranking is the second stage of two-stage retrieval. The first stage (BM25 + embeddings) fetches ~50 candidates; a dedicated reranker model (Cohere Rerank, Voyage Rerank, BGE Reranker) then scores each (query, document) pair more precisely and picks the top 5 to actually feed the LLM. Reranking dramatically improves citation quality but is expensive — only the top tier of engines run a reranker per query.
Example
Perplexity's pipeline: BM25 + dense embedding retrieve 50 → cross-encoder reranker scores pairs → top 5 by reranker score → LLM synthesizes from those 5.How to optimize
Rerankers reward direct topical match. A page tightly focused on the seed query outscores a tangentially-related authority page. Match query intent precisely on each landing page.