Signed up for the trial in one of my tenants today to play around with the studio. I probably have not configured something correctly, but I cannot seem to find out what to change in order to get reference to be correct. The response is right, but the reference is obviously wrong. I will read through the docs soon.
We are experiencing this issue as well. The answer is somewhat correct except one of the programs mentioned is not, but then the citation has nothing to do with the text referenced in the chat.
It has been over a month and we are still experiencing this exact issue. Is this going to get fixed?
There are a couple things that can be done that I have found. One could be to specify keywords within triggers and provide generative nodes to those which eliminates the fallback error I initially posted. Another interesting thing is that specific nodes can be set to omit specific document sources within their yaml as explained here:
https://learn.microsoft.com/en-us/microsoft-copilot-studio/nlu-documents#disable-documents-as-a-data-source-for-a-generative-answers-node
Until those improvements release (which, by the way, are we able to see when those are coming out?), is there anything we can do ourselves to improve the reference accuracy?
There are a couple of improvements being released in the next couple of weeks that should improve hallucination and citation issues.
I experienced the same thing: in many cases answer is ok but references are off-topic
Hello, I have the exact same issue. Initially, generated answer is OK, references to documents are OK just after setting the data sources to one of my folders. When asking the exact same questions a few hours later, the text is a bit different but still OK, but document references are totally wrong. As I asked to cite sources, wrong documents are cited also in the answer. I did not change anything in the indexed folder. By wrong document, I mean that the content of the proposed document is not at all related to the question nor the answer given, while the answer is still relevant. Looks like something gets broken in the index.
WarrenBelz
146,745
Most Valuable Professional
RandyHayes
76,287
Super User 2024 Season 1
Pstork1
66,091
Most Valuable Professional