I noticed that many DS folk(including myself) were tackling challenges in the search space. Sharing some of my notes on it. Assumes you know the basic SBERT architecture. Do check it out!
Scann is another good alternative for FAISS. Preserves accuracy and gives good optimization on inference time. It takes magnitude and direction into account when learning quantization and is unique
Very neat and crisp explanation!
I loved how succinct this is. I had tried the same BM25+SBERT method which works well.
Another method I can think of is getting all embeddings by passing them through SBERT and using Annoy/Faiss.
I am also maintaining a repo for info rel to Semantic Search. https://github.com/Agrover112/awesome-semantic-search
Would be great to have you contributing this article or any papers to our repo.
Scann is another good alternative for FAISS. Preserves accuracy and gives good optimization on inference time. It takes magnitude and direction into account when learning quantization and is unique