Sentence Transformers

Sentence Transformers (bi-encoders) are covered fully in Embedding — sections on dense embeddings, bi-encoders vs cross-encoders, training objectives (MNR loss, contrastive learning), domain adaptation, and evaluation.

Key points for quick recall:

• Encode query and document independently → dot product for similarity
• Embeddings are pre-computed and stored in a vector index (ANN search)
• Trained with Multiple Negatives Ranking (MNR) loss — other documents in the batch serve as negatives for free
• Strong at semantic similarity; weak at exact keyword matching
• Use alongside BM25/SPLADE in hybrid retrieval to cover both failure modes