You don't need a vector database

More than anything else I thought that the vibe of using traditional IR techniques over the new-fangled dense embedding vector ones is interesting. In many systems it's likely that the dense vectors actually don't offer any additional value.

BM25 being used over SPLADE was a bit odd though. I was also mildly confused by their hybrid-search approach. I would have expected that they would do the more modern technique of performing a sparse vector (full-text) search and dense vector search in parallel then re-ranking with a cross-encoder NN instead of re-ranking w/ the dense vectors themselves.