Sentence Transformers Library Adds Multi-Vector Encoder Support

The Sentence Transformers Python library has released version 6.0, which introduces a new model type called MultiVectorEncoder. This update allows users to load and utilize ColBERT-style late interaction retrieval models directly through the library's existing API. The new encoder supports various checkpoints, including those from PyLate, Stanford-NLP ColBERT, and ColPali-family models for visual document retrieval.

Unlike traditional dense embedding models that compress text into a single vector, multi-vector models retain one vector per token. This approach uses the MaxSim operator to compare query and document tokens, which may improve retrieval quality for complex queries or out-of-domain data. While this method offers increased precision by preserving token-level information, it requires significantly more storage space for indexes compared to single-vector models. The update aims to streamline the integration of these models into search stacks, providing a unified interface for developers working with dense, sparse, and late-interaction retrieval architectures.