I asked this same question here and wasn't able to get an answer that worked. I tried a variety of instructions in the knowledge agent and personality agent. What I'm doing for now is: before the bot responds, putting the knowledge agent's answer into an AI task that says to remove that kind of language, and then having the bot respond with the "cleansed" answer. It has worked well so far. The drawbacks are that it adds expense and longer delay before the bot responds, but I have no choice, I can't have the bot giving hallucinated citations and links.