[name] asked: How did you consider the inference cost if instead of 2–3 relevant chunks, you could send the model a whole article?