Preserving quality while reducing vector counts
Hugging Face’s technical article dated November 11 brings pooling, which combines similar token vectors, into the training process. The aim is to reduce storage requirements for large document collections while preserving search quality.
In the reported experiment, reducing document vectors to one quarter of their original count retains 94.6% of the NanoBEIR nDCG@10 score with standard training. Pooling-aware training raises that figure to 97.8%.
A reproducible example
The article provides a Sentence Transformers training example and explains the method. The reported percentages apply to that experiment’s data and settings; gains on another search collection require separate measurement.