Can you give some examples of the low quality/incorrect responses? Then try remedying those by rephrasing the prompts? I’m curious what the actual limitations are.