Sentence Transformers v6.0 adds ColBERT-style multi-vector retrieval
Sentence Transformers v6.0 introduces MultiVectorEncoder, a fourth model type for ColBERT-style late-interaction retrieval. The Python library already supports dense, sparse and reranker models for uses including semantic search and retrieval-augmented generation. MultiVectorEncoder can load PyLate and Stanford-NLP ColBERT checkpoints, while ColPali-family models support visual document retrieval through the same API.
Unlike a conventional embedding model, which compresses a document into one vector, a multi-vector model retains one projected vector for each token. At scoring time, MaxSim finds the strongest document-token match for every query token and sums those similarities. This preserves token-level evidence, including exact identifiers and separate requirements in a query, but produces a substantially larger index.
In the article’s Natural Questions example, 4,874 passages generated 608,414 128-dimensional token vectors, occupying 311.5 MB in float32. A compressed fast-plaid index reduced that representation to 92 MB. The package can be installed with pip; Sentence Transformers v6.0 requires transformers v5.x, torch 2.2+ and huggingface-hub v1.x. Image dependencies are additionally required for ColPali-style visual retrieval.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.