問題文
A bank wants a chat assistant that answers using its own internal policy documents, which the model was never trained on. The team is told the model weights will not change at all. How can the answers still rely on those documents?
選択肢
- The stricter guard that can block the unsupported assertions, and the fact that weights stay untouched settles it.
- The nightly sweep that can memorize the visited directories, and the fact that weights stay untouched settles it.
- The fetched passages that can enter the assembled prompts, and the fact that weights stay untouched settles it.
- The uploaded dump that can overwrite the pretrained layers, and the fact that weights stay untouched settles it.