Bi-Encoder

A retrieval architecture that encodes the query and document separately into fixed-size vectors and computes relevance with a simple function such as a dot product or cosine similarity. Document vectors can be precomputed and searched with approximate nearest-neighbor methods, enabling low-latency candidate retrieval over large collections.

Separate encoding reduces fine-grained interaction between query and document tokens. A bi-encoder therefore often serves as the first pipeline stage: it maximizes recall within the budget and hands a shortlist to reranking. Quality depends on embedding training, domain fit, negative examples, and index configuration; speed alone does not show that the required evidence was retrieved.