What is Reranking?
Also called Semantic reranking
Definition
Reranking is a later retrieval stage that reorders an initial set of search results using another relevance assessment. It can improve which candidate documents reach an answer generator without changing the original document collection.
The operation of Reranking is explained.
A fast retriever first finds candidates. A reranker then evaluates a smaller set, often with a more computationally intensive model. Elastic documents semantic reranking as applicable to lexical, semantic, and hybrid retrieval.
The relevance of Reranking is explained.
A SaaS support assistant might retrieve several access-control articles, then prioritize the one addressing the requested identity provider.
Reordering candidates is useful only if the needed article was retrieved in the first place.
The evaluation of Reranking is explained.
Retrieval evaluation measures candidate coverage and final ranking separately. Labeled questions support comparisons of relevance, latency, and failures. A reranker cannot recover a document absent from its candidates.
Frequently asked questions about Reranking
Is reranking initial retrieval?
Reranking follows an initial retrieval stage and reorders its candidates.
Can reranking work with keyword search?
Semantic rerankers can reorder candidates produced by keyword search.
Does every answer engine expose its reranker?
Public answer engines do not necessarily disclose their internal reranking implementation.