Comment by xfalcox

Also worth mentioning that we use quantization extensively:

- halfvec (16bit float) for storage - bit (binary vectors) for indexes

Which makes the storage cost and on-going performance good enough that we could enable this in all our hosting.