Sentence Transformers v6.0 Adds Multi-Vector Search Support

Sentence Transformers v6.0 introduces MultiVectorEncoder as its fourth model type, supporting ColBERT-style Late Interaction search through the standard API. Existing PyLate and Stanford-NLP ColBERT checkpoints can be used directly, and colpali-engine models used for page image search also work with the same API.

While standard Dense embedding models compress sentences into a single vector, multi-vector models preserve token-level vectors to improve search quality. This improves accuracy for queries containing rare proper nouns or complex conditions, but at the cost of larger index sizes. For example, encoding 4,874 Natural Questions paragraphs generates an index approximately 42 times larger than that of a dense model; however, applying compression techniques such as fast-plaid allows it to be managed at a capacity similar to existing dense indexes.

This update unifies LightOn's PyLate ecosystem with Sentence Transformers, enabling developers to build high-performance search systems without learning a separate library.