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

Sentence Transformers library version 6.0 introduces a new model type called MultiVectorEncoder, which implements ColBERT-style late interaction retrieval.

This allows token-level matching using the MaxSim operator, preserving more information than dense single-vector embeddings, particularly beneficial for multi-requirement queries and visual document retrieval.

The update enables loading of various checkpoint formats including PyLate, Stanford-NLP ColBERT, and colpali-engine models through the same API.