The limitations of RAM-based indexes for approximate nearest neighbor search are storage size constraints, mainly due to auxiliary data like vector ids and links. Lossy vector compression techniques have been used to reduce index size, but new lossless compression schemes like asymmetric numeral systems and wavelet trees can compress vector ids by a factor of 7, reducing index size by 30% on billion-scale datasets without affecting accuracy or search speed. Surprisingly, these methods can also compress quantized vector codes losslessly by exploiting flaws in the original quantization algorithm. The source code for this approach is available online.
https://arxiv.org/abs/2501.10479