GPT4 hooked up to a legal DB and instructed that if no knowledge from the DB can be referenced to respond negatively would've solved this completely.
Even a hard link to rewrite the prompt from "use provided knowledge to answer the question" to "no knowledge could be found, inform the user of this fact in a way that keeps the conversation fluid. Recommend that they search other sources themselves." would be more reliable.
The problem, as always, isn't the tool. It's the tool using the tool.
Even a hard link to rewrite the prompt from "use provided knowledge to answer the question" to "no knowledge could be found, inform the user of this fact in a way that keeps the conversation fluid. Recommend that they search other sources themselves." would be more reliable.
The problem, as always, isn't the tool. It's the tool using the tool.