Power Platform Community Forum Thread Details

I have developed many copilot agents using copilot studio, which have SharePoint sites and datavesrse tables as their source.. but i have those questions that i can not find a clear answers on them:-

1- Is there an order to how knowledge sources are indexed by a copilot agent? and if the source is SharePoint does Copilot use the classic SharePoint search ?

2- Why do we sometimes get inconsistent answers when asking the same questions?

3-Is there a way to integrate a PowerBI list instead of a SharePoint list?

4- Can the agent learn through user interaction? does generative AI need to be turned on for this to occur? In other words, if we have the same source, should we expect the same Copilot agent to provide more clear answers in the future, due to improvement to the underlying AI algorithms?

Thanks

Categories:

General topics

Hey John,

my understanding is:

1. A copilot agent doesn't actually index content sources; it queries them and then probably uses some form of ranking on the results. And no, it's not using classic SharePoint search, it is using SharePoint's Semantic search, via an internal (to Microsoft) Graph API.

2. It's AI and currently most LLMs won't provide consistent results, especically if there is any contention/conflicting information in the source content.

3. When you say a PowerBI list, do you mean a table in a Power BI workspace? If you do, then kinda....

I believe there's a connector for Fabric Data Agents, which includes support for Power BI semantic models, Warehouses, Lakehouses and KQL databases, which might provide what you're after.

4. Copilot Studio agents only remember information from the current session, there's no real persisted learning. Though it's plausible that you could build somethign to monitor conversations, identify new/correct knowledge and add that to a document in SharePoint, which could then be specifically included in future sessions. A nice little project for someone...

Hopefully that's helpful.

If you found out anything else that's useful or more accurate, please do update this post.

Thanks

Craig