Sentence Transformers 6.0 adds native ColBERT late-interaction search
MultiVectorEncoder brings PyLate, Stanford ColBERT and ColPali image retrieval to one API — better recall, ~42x storage. Plus OpenAI eyes ZDR API traffic.
Copy markdown
One API, four model families
Sentence Transformers v6.0 adds MultiVectorEncoder, a ColBERT-style late-interaction family beside dense, sparse, and rerankers. It loads PyLate, Stanford-NLP ColBERT, and ColPali checkpoints from the Hub, scores with MaxSim, and plugs into Qdrant, Weaviate, Vespa, and LanceDB. On NanoBEIR, the 149M LateOn hit 0.6868 NDCG@10, beating its dense sibling DenseOn (0.6764) on 9 of 13 datasets.
Search page images without OCR
ColPali support lets you embed document screenshots and PDFs as multi-vector representations and retrieve over them directly — a weekend build for RAG over scanned docs with no OCR step, using the same encode-retrieve-rerank pipeline as text.
The catch: ~42x more vectors to store
One vector per token, not per document: 4,874 Natural Questions passages grew to 311.5 MB in float32 versus 7.5 MB dense — clawed back to 92 MB with PLAID indexing. Exhaustive MaxSim ran 98 ms on GPU; indexed queries returned in 11-18 ms. Budget the storage before you migrate.
Elsewhere: OpenAI will scan even ZDR API traffic
OpenAI's new Private Safety Processing, rolling out in September, runs automated misuse detection across API interactions — including zero-data-retention accounts — returning limited safety signals without exposing prompts to staff. If you lean on ZDR for sensitive data, your traffic is now in scope.
Elsewhere: GitHub revokes credentials by token type
GitHub now lets you revoke or deauthorize credentials scoped to a single token type — OAuth, PAT, or app — instead of nuking everything at once. Faster, surgical cleanup the next time a CI secret leaks.