after completion of #100 we have a nice dataset of 64k diverse questions. we can select most popular ones from it, and "benchmark" documentation with AI agents by asking them to find answers in docs. this will show us what topics are being asked but not being covered in docs clearly.