How IndexRefineFlat gets the real L2 distance?

[ghrua]

Hi, I am a beginner of FAISS. This is a really powerful toolkit! I am now curious about a feature of FAISS when reading the wiki.

In the wiki about IndexRefineFlat, the author said that IndexRefineFlat is able to re-rank the search results with real distance computations. However, when computing the real distance, the information of the original vectors, rather than the vectors encoded by quantizers, is necessary.

I am curious about how IndexRefineFlat knows the original vectors of search results, especially when some lossy quantizers, e.g., PQ, are used. I think the information of original vectors will be dropped when the index is built?

Thanks for your kind explanation :)