I'd like to benefit from the reduced disk footprint with RaBitQ quantization such as Q4 and Q8 in exchange for some recall accuracy drop. Having smaller footprint would allow for faster user adoption.
Note RaBitQ with 1 bit quantization can be used for index search candidates, and then with full precision among tye candidates, this significantly speeds up recall without loosing accuracy.
https://lancedb.com/blog/feature-rabitq-quantization/
https://docs.vectorchord.ai/vectorchord/usage/quantization-types.html#quantize-to-rabitq8-vector
I'd like to benefit from the reduced disk footprint with RaBitQ quantization such as Q4 and Q8 in exchange for some recall accuracy drop. Having smaller footprint would allow for faster user adoption.
Note RaBitQ with 1 bit quantization can be used for index search candidates, and then with full precision among tye candidates, this significantly speeds up recall without loosing accuracy.
https://lancedb.com/blog/feature-rabitq-quantization/
https://docs.vectorchord.ai/vectorchord/usage/quantization-types.html#quantize-to-rabitq8-vector