Announcements
Hello! We recently defined a data source, in a Word document, but when questions are asked about topics NOT in that document, the AI makes up answers, which we don't want to happen. I tried changing the moderation to "high" but the same situation still occurs.Do you know any other alternative that could help me?thank you so much
Can you open a support request so this can be investigated further?Get Help + Support in Power Platform - Power Platform | Microsoft Learn
Please provide these as part of the case description:
Tenant ID:
Environment ID:
Bot ID:
Conversation ID: (type "/debug conversationid" in chat canvas)
Use the thumb down button: In the test pane, click on the thumb down button shown when Copilot Studio returns an inconsistent answer so that contextual data on the inconsistent answers are recorded.
botContent.zip snapshot: Please attach the botContent.zip file for the conversation where you experienced the issue (from the test pane, click ... and "save snapshot")
User utterance:
Actual generated answer:
Expected generated answer (or explanation of what is wrong with the generated answer):
Generative Answer source(s) (documents and/or URLs):
Please paste or add screenshots as attachments.
Hi, did you find any solution on this?
Good morning,From what I understand, changing moderation blocks toxic answers but doesn't check that the answers are correct. If you want to keep only truthful answers, you should check the temperature and the top_p parameters.A combination that worked for getting truthful yet entertaining answers was temperature = 0.1 and top_p = 0.4Have a great day !
Good morning!!
Thanks for your reply, I was thinking of changing the same things but where can we change that? Do we have option in the copilot studio to change those things? It would be helpful if I got that.
For azure openai, it is in the connection properties. Other types of generative answers rely on microsoft copilot to extract the information from the document so I'm not sure you can modify directly the parameters of the model.
ya, thanks
Under review
Thank you for your reply! To ensure a great experience for everyone, your content is awaiting approval by our Community Managers. Please check back later.
Congratulations to our 2026 Super Users!
Congratulations to our 2025 community superstars!
These are the community rock stars!
Stay up to date on forum activity by subscribing.
Valantis 279
Romain The Low-Code... 222 Super User 2026 Season 1
Arild Aarnes 64 Super User 2026 Season 1
