Sentence Transformers v6.0 Introduces Native Support for Multi-Vector Late Interaction Models
Hugging Face has launched version 6.0 of its Sentence Transformers library, which introduces support for multi-vector embedding models that previously depended on LightOn's PyLate library. These multi-vector models, unlike dense models, keep a vector for each token, enhancing retrieval accuracy for complex queries and lengthy documents, albeit with an increased index size. Using the LateOn model, 4,874 passages from Natural Questions were encoded, resulting in 608,414 token vectors. The library is compatible with checkpoints from PyLate, Stanford-NLP ColBERT, and ColPali models. Multi-vector models are applicable for exhaustive searches, reranking, or can be indexed in various databases. This release is attributed to Omar Khattab, Matei Zaharia, Antoine Chaffin, and others.
Key facts
- Sentence Transformers v6.0 adds native support for multi-vector (late-interaction) embedding models.
- Multi-vector models keep one vector per token and use the MaxSim operator for scoring.
- MaxSim computes the sum of each query token's maximum similarity against document tokens.
- Encoding 4,874 Natural Questions passages with LateOn produced 608,414 token vectors, about 42x the storage of a MiniLM dense index.
- The library loads checkpoints from PyLate, Stanford-NLP ColBERT, and ColPali-family visual document retrieval models.
- ColPali-style models enable visual document retrieval with text queries against page images, no OCR.
- Multi-vector indexing is supported natively in Qdrant, Weaviate, Vespa, LanceDB, VectorChord, Milvus, and fast-plaid.
- Sentence Transformers v6.0 requires transformers v5.x, torch 2.2+, and huggingface-hub v1.x.
Entities
Artists
- Omar Khattab
- Matei Zaharia
- Antoine Chaffin
- Raphael Sourty
- Paulo Moura
- Amélie Chatelain
- Manuel Faysse
- Hugues Sibille
- Tony Wu
- Bilel Omrani
- Gautier Viaud
- Céline Hudelot
- Pierre Colombo
- Benjamin Clavié
- Griffin Adams
- Kenneth Enevoldsen
- Roman Solomatin
Institutions
- Hugging Face
- Sentence Transformers
- LightOn
- Stanford NLP
- PyLate
- colpali-engine
- fast-plaid
- Qdrant
- Weaviate
- Vespa
- LanceDB
- VectorChord
- Milvus
- OpenSearch
- Elasticsearch
- turbopuffer
- MTEB
Locations
- Antarctica