Bi-Encoder
Bi-Encoder
A bi-encoder (dual encoder / two-tower model) encodes queries and documents independently into dense vectors, then computes relevance via a simple similarity function (dot product or cosine). Documents can be pre-encoded offline.
Query → [Encoder_Q] → q ∈ ℝ^d
→ sim(q, d) = q · d
Document → [Encoder_D] → d ∈ ℝ^d
Key advantage: document vectors are computed once and indexed → retrieval is just nearest neighbor search in vector space.
See Dense Retrieval for full details, training, and indexing.