Sentence Transformers v6.0 Adds Multi-Vector Encoder Support

The Sentence Transformers Python library has released its v6.0 update, which introduces the MultiVectorEncoder. This new model type supports ColBERT-style late interaction retrieval, a method that preserves fine-grained token-level signals by scoring queries against documents using the MaxSim operator. This approach aims to improve retrieval performance compared to traditional dense embedding models that compress text into a single vector.

Users can now finetune existing multi-vector models or build new ones from base transformers. The update allows for custom configurations, such as adjusting document length limits and implementing punctuation skiplists to optimize indexing. According to the documentation, domain-specific finetuning can help models adapt to specialized vocabularies and document lengths that general-purpose retrievers may not handle effectively. The library provides a structured approach to training, including components for datasets, loss functions, and evaluators, enabling developers to build and refine models on their own data.