Bi-Encoder
A retrieval architecture that independently encodes queries and candidates into vectors so similarity can be computed efficiently at large scale.
Information-Retrieval Context
A bi-encoder maps a query and each candidate independently into vector representations. Candidate embeddings can therefore be precomputed and indexed for low-latency ANN retrieval over a large corpus.
Ranking Boundary
Independent encoding compresses query-document interaction into vector similarity, which is cheaper but less expressive than joint encoding. Domain shift, training objectives, and hard-negative selection can materially change recall and precision.