Power Platform Community Forum Thread Details

We have been trouble shooting a CopilotStudio setup that originally worked flawless. The first solution used a separate setup that we re-created in a new location.

Both the old and the new solution builds on a GPT35turbo deployment an Cognitive search and a Storage account with a chat container deployed.

Custom data was first uploaded through the Azure Open Ai studio Playground -> Add data -> Upload data. It successfully uploaded the data to the SA and created a chunks container. An index and indexer was added to the Search instance and it all worked good in the Playground. We deployed to a Copilot, changed the url to the specific custom Environment and created the Azure OpenAi connection. Verified it in make.powerapps.com/connections.

After that we created the Copilot manually, setup the Generative Ai option, adjusted the Conversation boosting and Generative text by selecting the AOI connection option, added Connections properties with same Deployment as in Playground, API version: 2023-08-01-preview, 800 tokens, 0 temp, 1 Top P . Added model data:

But this time it only throws blank answers and automatically Fallsback with the dummy: I’m sorry, I’m not sure how to help with that. Can you try rephrasing?

We have verified that the new setup works outside CopilotStudio, but with app deployment and api calls. We have tried removing and renewing the AOI connections in Power platform but without success and now would appreciate any guidance...

Categories:

General topics

This took way to long and is still kind of on-going. But we are now 99% positive that this is a time out issue. After a lot of support and meetings with Microsoft engineers across all platforms, the Azure specialists could provide us with some logging for what was hitting the Cognitive search. Logs showed us a lot of bad requests but could not support on why, since the call looks correct. Later down the lane the Power platform engineers reported back that they now could see Client time outs. By this time the client and we had already moved on to use Power automate flows calling the AOI and data instead of the OOTB MCS system topic solution, but in the flows we eventually also got time out issues. However PA uses a 120 second threshold and 8 retriest in HTTP action per default, so this was almost never an issue.

But after receiving this new time out log finding from the PP engineers, we switched back to GPT3.5-turbo and limited the token query and all of a sudden it started working again, hitting back accurate and nicely formated responses instead of a the default Fallback topic...

It turns out MCS uses a 30 second timeout and no retries in the call, which is currently way to limited and will then result in a "no match" fallback response...

Best of luck!