Bi-encoder
Neural network architecture that encodes query and document independently to vectors, then compares with dot-product/cosine-sim. Fast (precompute doc vectors offline) but less accurate than cross-encoder.
Foundation for modern vector search + RAG: sentence-transformers, OpenAI text-embedding-3, BGE, Cohere Embed, Voyage. Train via contrastive learning (positive vs negative pairs). Embedding dimension typically 384-1536. Use: index 1M docs once → query-time only encode query (5-10 ms) + vector search. Distinct from cross-encoder: bi-encoder optimized for retrieval, cross for re-ranking. Standard hybrid approach: bi-encoder for broad retrieval (top-100), cross-encoder for precision (top-10).