Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers

Multi-vector (late interaction) models are now supported in the Sentence Transformers library, offering an alternative to standard dense embedding models. Unlike dense models that compress text into a single vector, these ColBERT-style models retain a separate vector for each token. Scoring is performed using the MaxSim operator, which calculates the maximum similarity for each query token against all document tokens and then sums these maximums. This method preserves token-level information, leading to stronger retrieval performance for applications like semantic search and visual document retrieval, though it requires a larger index.