Asking things like "Who voted for me the most" won't be very accurate to ask an LLM, because the output is generated, not pulled out, so the LLM might decide not use the most probably data, but a least probable one, which will give you bad results for questions that needs pulling data instead of generating it.