You could just have a specialized model that rewrites prose questions to actual search keywords (i.e. what power users do in their heads), and even a 2nd one that does a fitting pass based on intention after querying.

Answering with AI slop is a choice.