I agree with you, but the question is more whether existing LLMs have enough training with AST queries to be more effective with that approach. It’s not like LLMs were designed to be precise in the first place