Vector databases have been indexing embeddings at full precision for no good reason, according to new research.
A team behind the open-source SuperKMeans project revisited three standard vector-search techniques: dimensionality reduction, quantization, and dimension pruning. They applied each one before clustering instead of after, then measured how it affected storage, clustering time, and centroid quality. The headline result: compressing embeddings down to 1-bit codes still produces clustering quality within 1% of full-precision vectors. That same compression cuts storage requirements by 60x and speeds up clustering along the way.
Vector search underpins semantic search and retrieval-augmented generation pipelines, and storage plus clustering time are two of the biggest costs at scale. If 1-bit codes really do hold onto near-optimal centroid quality, teams running large embedding indexes could shrink infrastructure spend substantially without touching their underlying models.
The paper doesn't cover every embedding model or dataset in production use, so the 60x figure is a promising benchmark result for now, not a blanket guarantee.