Currently knnPerfTest.py can test int4 and int7 KNN quantization (and float32, no quantization). Lucene also supports "simple binary quantization", though it's not so turnkey because the user must pre-quantize their vectors to single bit per dimension in byte[] vector form.
OpenSearch has such a binary quantizer to run outside of Lucene -- maybe we can poach that here? Or maybe improve Lucene so it can do this quantizing under the hood?
Currently
knnPerfTest.pycan testint4andint7KNN quantization (andfloat32, no quantization). Lucene also supports "simple binary quantization", though it's not so turnkey because the user must pre-quantize their vectors to single bit per dimension inbyte[]vector form.OpenSearch has such a binary quantizer to run outside of Lucene -- maybe we can poach that here? Or maybe improve Lucene so it can do this quantizing under the hood?
Lucene now also (soon?) has RabitQ inspired "better binary quantizer" which we should also add to
knnPerfTest! So hard to keep up...