Agreed on both counts. I do the same thing with headings and I use the results of the retrieval step to display a list of relevant docs while the answer is generating.

The latest models are way better and faster than GPT-4 was. You’ll probably be happy when you get back into it.