The radix selection algorithm is a modified radix sort that, since we don't care to sort the whole list but rather just the top K results, manages to cut some of the time out of the algorihm. It was an awesome way to get experience in CUDA programming.

Radix sort works with any ordered sets--it just lends itself well for GPUs since it is designed to be run in parallel. I used the modified version to get the best hamming distance results quickly, and implemented a few other distance measures as well (e.g., cosine distance)

You can quantize an embedding: https://zilliz.com/learn/what-are-binary-vector-embedding#Ho...

This is a great question! Yes, for a very specific subset of problems--those where you need total recall. HNSW-based algorithms typically only compare against a subset of the whole collection in order to achieve much faster results than linear time search, and it is sometimes the case that they miss the true best results, which is a trade-off worth making. I aimed to keep total recall and my method does in fact perform faster for this particular use case.