RaBitQ
A quantization compression method for high-dimensional vectors. It applies a random rotation or orthogonal transformation and then represents each vector with about one bit per dimension while retaining a computable error upper bound. It is suitable for two-stage coarse filtering followed by fine ranking in retrieval.
Timeline (last 90 days)
PRISM is a native Postgres index access method that requires no additional deployment of a separate service; it clusters vectors by k-means, stores them using RaBitQ encoding, and during queries first walks the centroid tree to find the nearest cluster, scans the posting list, filters using error bounds, and then reranks with full-precision vectors.
pg_vectorsearch uses RaBitQ encoding, storing each vector with an encoding of about one bit per dimension, thereby supporting a memory-resident index at billion scale.
All Facts (2)
PRISM is a native Postgres index access method that requires no additional deployment of a separate service; it clusters vectors by k-means, stores them using RaBitQ encoding, and during queries first walks the centroid tree to find the nearest cluster, scans the posting list, filters using error bounds, and then reranks with full-precision vectors.
50%Unverifiedpg_vectorsearch uses RaBitQ encoding, storing each vector with an encoding of about one bit per dimension, thereby supporting a memory-resident index at billion scale.
50%