"Thinking" models do that, now. They either search for the context (token usage goes brrr...) or ask for more information before going further.