For your use-case, instead of using RAG, I would recommend having the LLM generate Typesense

filter_by

queries for you, so you don't even have to use embeddings inside of Typesense. That will produce much better results. This article describes the concept: https://typesense.org/docs/guide/natural-language-search.html We're using Gemini in that example, but the same concept can be used with any LLM.