The paper is a bit odd. It feels like a lot of things are missing: - How are the quantized vectors used to do NNS? Do we have to brute force compare against ea

The paper is a bit odd. It feels like a lot of things are missing: - How are the quantized vectors used to do NNS? Do we have to brute force compare against each codeword? - In thst case, how many codewords are there? How big is k compared to n? How does changing k affect performance? - What architecture is used to train the embeddings? We are only given the loss function. - How fast is the preprocessing step? If we need to do Lloyd's algorithm type iterations combined with training neural nets, this is potentially a lot slower than Faiss which just picks random landmark points.