Sentence Transformers v6.0 adds MultiVectorEncoder for ColBERT-style retrieval
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Multi-vector models now store one embedding per token instead of one per document, boosting retrieval accuracy by 10–20% in benchmarks. This means your RAG pipeline can finally match rare terms or exact phrases without OCR, but your index size and query latency will grow 10–100×—plan for 100 GB RAM per million docs and sub-second response only with GPU-accelerated MaxSim.
Sentence Transformers v6.0 now natively supports ColBERT-style late interaction and ColPali visual document retrieval checkpoints through its standard Python library API. For production RAG and search pipelines, this allows you to completely bypass OCR bottlenecks by matching text queries directly to document page images, though it requires managing significantly larger vector indexes since the model stores one vector per token instead of a single averaged embedding.