I’m finding smaller vector databases can be almost ephemeral if you avoid parsing : https://hushh-labs.github.io/hushh-labs-blog/posts/you_dont_need_a_vector_da

I’m finding smaller vector databases can be almost ephemeral if you avoid parsing : https://hushh-labs.github.io/hushh-labs-blog/posts/you_dont_need_a_vector_database.html https://hushh-labs.github.io/hushh-labs-blog/posts/you_dont_... I can retrieve a query, encode the embedding, load the vector store, calculate KNN, and return rendered results in under 50 milliseconds. It’s even faster if you simply cache the vector store. Really interested in this space and hoping to hear more ideas.