Is LLM a must in Retrieval augmented generation?

Sorry if its a dump question. So, I have been playing around with RAG to make sense of domain specific data. I see that the common practice is to use the embedding search result with LLM. My question is, do we really need LLM here? Of course, the output is not very elaborative.
(I just tried one of the pdf chat tools. I asked a question about something present in the pdf. The question was about a specific number in the pdf. The answer started with the value. And it then added the definition of the topic in the question. This came from LLM. So, if one is not interested in that, wouldn’t using RAG alone not suffice?)

Welcome to the forum, kprmlr!

I’m RAG stands for Retrieval Augmented Generation. So RAG specifically is for LLM’s.

I believe the actual question you wanted to ask is: can we use vector database search results without LLMs. Yes, we can, and in some cases it makes sense.

This is not the case where you can use vector database alone, as it only gives you a full chunk of data stored there (which I doubt is just the value you are looking for; it would be an abstract or a full page), so in this case both came as 1 message from LLM, which was fed a relevant chunk from the vector database.

1 Like