Reranking
also called Reranker, Cross-encoder
A second pass that reorders candidates by actual relevance.
A model that scores each candidate against the query directly and reorders the results, after a fast first-stage retrieval has narrowed the field.
Retrieve loosely, rerank strictly. Fetch a hundred candidates cheaply, score them properly, keep the top five. This small step often improves answer quality more than swapping the generation model.
The pattern is a funnel: cheap retrieval casts a wide net, expensive scoring orders the shortlist. It works because a cross-encoder that reads query and candidate together is much more accurate than comparing two independently-computed vectors — and because you only pay that cost on a few dozen candidates.
The highest-leverage step in most RAG systems
The right document retrieves but sits at rank eleven, outside the top ten.
Reranking pulls it into the top three, so it survives the context budget. Where the budget only fits five chunks, the difference between rank eleven and rank two is the difference between a correct answer and a confident wrong one.
- blogContextual retrieval in AI systems ↗
The practical hybrid-retrieval article: chunking, embeddings, BM25, rank fusion and reranking, plus prepending chunk context to reduce retrieval failures.
- paperColBERT: efficient and effective passage search via contextualized late interaction over BERT ↗
Late interaction: encode query and documents separately, then compare token-level representations at retrieval time. The accuracy/latency middle ground in neural ranking.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Semantic Search
Search by meaning instead of matching words.
Hybrid Search
Keyword search and vector search, combined and merged.
Evaluation · Evals
How you know the system is actually working.