フリー問題

問題文

Which end-to-end evaluation confirms that a retrieval and generation solution works as intended?

選択肢

1. Measuring both whether the correct passages are retrieved and whether the final answers are grounded and relevant, on the same set of questions
2. Measuring the retrieval quality only, since it determines everything downstream and a generation step cannot invent an answer that the passages do not support
3. Measuring latency and token consumption only, since those are objective numbers while any quality score depends on a judgment that can be disputed
4. Measuring the final answer quality only, since that is what users see and a good answer means every layer beneath it must have worked correctly