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.

The other way round. You stick a search engine in front of your LLM to selectively feed it facts relevant to your query, which the LLM synthesizes in its own words.

To oversimplify, its more like sending a block of text to an LLM and asking it to answer a query based on that block of text.