問題文
An assistant must not reveal other customers' data, must stay on the company's topics, must not produce harmful content, and must not perform actions above a value limit. Where does each control go?
選択肢
- A single output check for all four, since every requirement ultimately concerns what reaches the user and inspecting the final response therefore covers all of them at once.
- Retrieval filtering and output checks for the data, input and dialog checks for the topics, input and output checks for harmful content, and application-side validation for the actions.
- A strict system-level instruction for all four, since stating the constraints clearly is what makes the model comply and additional layers only add latency and false blocks.
- Input checks for all four, since preventing a request from reaching the model is the most reliable way of ensuring none of the four outcomes can occur.