Hugging Face has published guidance on using multi-vector, late-interaction embedding models with Sentence Transformers. The approach is designed for retrieval systems that need more detail than a single vector can preserve.
Traditional embedding search often compresses a document or passage into one vector, a list of numbers that represents meaning for similarity search. Late-interaction models keep multiple vectors and compare them later in the retrieval process, which can better capture which parts of a query match which parts of a document.
The trade-off is complexity. Multi-vector retrieval can improve ranking quality in some systems, but it usually needs more storage and a more careful search pipeline than single-vector embeddings.
For developers building retrieval-augmented generation, the useful point is not that one method replaces the other. It is that Sentence Transformers now gives a clearer path to experimenting with richer retrieval when simple embedding search misses important matches.