That vector sets design outlined towards the end of the article is delightful - it's exactly the kind of API design I appreciate from Redis over the years: simple, elegant and feels obviously correct to me.
Where would a Redis vector store play a part though? Maybe you'd load up relevant embeddings for a particular user while they're interacting with their dataset, to make their responses quicker? You've already spent the effort on hydrating their data out of persistence though. I guess step one is likely being a more trusted alternative to the in-memory vector solutions like HNSW, Faiss, and a potentially faster engine than pg_vector. I've always seen Redis as an augmentation, but maybe in this role it can take the helm?
It's exactly that. Redis is an in-memory data structure server that you can outsource index-style operations to. Vector similarity is a type of index search. I think it's an exact fit for Redis.
