Comparison
Bi-encoder vs cross-encodervsReranking
Bi-encoder vs cross-encoder
you embed queries and documents separately so the index can be built ahead of time, then let a slower model read the pair together for the few you actually rank.
A bi-encoder embeds query and document separately, enabling precomputed indexes; a cross-encoder reads them together and scores the pair, more accurate but unindexable. Retrieval uses bi-encoders, reranking uses cross-encoders. That sentence explains the whole retrieval-then-rerank architecture.
Full entry →Reranking
you fetch fifty candidates cheaply and then have a slower, better model reorder them before you show the top five.
A second pass: retrieve a generous candidate set (say 50) with fast search, then rescore the top of it with a slower, more accurate model and keep the best few. Cheap first pass for recall, expensive second pass for precision.
Full entry →