I'll also attempt an answer. You chop up your documents into chunks. You create fancy numbers for those chunks. You take the user's question and find the

I'll also attempt an answer. You chop up your documents into chunks. You create fancy numbers for those chunks. You take the user's question and find the chunks that kind of match it. You pass the user's question with the document to an LLM and tell it to produce a nice looking answer. So it's a fancy way of getting an LLM to produce a natural looking answer over your chunky choppy search.