Hugging Face explains pooling-aware training for compressed multi-vector retrieval
AIHugging Face explains how token pooling shrinks multi-vector embedding stores and how pooling-aware training preserves retrieval quality at the same vector budget. The post says a normally trained model retains 94.6% of unpooled NanoBEIR nDCG@10 at pool_factor=4, while pooling-aware training raises retention to 97.8%.
Source details
Hugging Face Blog · huggingface.coPublished · added here
Original title: Token Pooling-Aware Training for Multi-Vector Embedding Models