We built a vector search engine that lets you choose precision at query time

ClickHouse introduced QBit, a new column type that stores floating-point numbers as separate bit planes, allowing users to choose precision levels during vector search queries rather than beforehand. This flexible approach lets developers dynamically balance search speed and accuracy without storing duplicate data or making irreversible precision decisions. Benchmarks showed nearly 2x speed improvements while maintaining good recall accuracy, with even heavily reduced 5-bit precision producing surprisingly meaningful results for semantic searches.