Hugging Face blog explains token pooling for multi-vector embeddings

Hugging Face published a technical blog post on November 11 explaining token pooling for multi-vector embedding systems such as ColBERT. The post says pooling can merge similar token embeddings to cut document vectors by 4x while keeping retrieval quality close to baseline. In one benchmark, a normally trained model retained 94.6% of its unpooled NanoBEIR nDCG@10 at `pool_factor=4`, while pooling-aware training reached 97.8% at the same storage budget. The article also outlines implementation controls, calibration, and a new `MultiVectorTokenPoolingLoss` wrapper in Sentence Transformers.