Scann vs faiss about retro-pytorch HOT 6 OPEN

scann is opensource but not really packaged well + it doesn't quantize

but also it's not really a blocker either way in the current state of this repo
doing more experiments with the LM and the LM+knn integration is

from retro-pytorch.

I agree is a bit cumbersome to use, but it should have (rather efficient) quantization, no? See last section of https://medium.com/@kumon/similarity-search-scann-and-4-bit-pq-ab98766b32bd

from retro-pytorch.

scann is fast but no it doesn't optimize for memory use, since it uses PQ4 it requires storing the embeddings at full precision for reranking to avoid loss of recall

from retro-pytorch.

@rom1504 thanks for the answer! So is it correct to say that the open source version of scann does use quantization to compute faster inner products? I mean there's implemented options for using brute force and 2 different quantizers (lut16 and lut256), but I see your point about storing the embeddings. That's also odd since reorder (which I assume is what you mean by reranking) is optional

from retro-pytorch.

Yes it uses quantization to compute faster inner product.
Do they report good results on pq4 and no reordering?

from retro-pytorch.

https://github.com/erikbern/ann-benchmarks/
https://ann-benchmarks.com/glove-100-angular_10_angular.html

from retro-pytorch.