So Generative Answers is one of the newer, and more powerful, features that Co-pilot Studio now has, and it can be found kind of across the entire Studio in little areas, but the main point is it’s allowed us to take in any sort of data source — from public websites to SharePoint and whatever we feed it — and it can answer questions that the user provides. No need to do any sort of extra work on extracting the data manually. Just feed it the question, and it outputs the answer based on the data sources hooked up.
And you can add this node to the topics that are already used. As can be seen in the photo there, it’s simply “Create Generative Answers.” But if a whole new topic is not desired, the out-of-the-box Conversational Boosting that is built in when Generative Answers is enabled can be used, and that topic will run whenever there’s some sort of unknown intent from the user. All that is needed is to supply the links to public websites or SharePoint sites, and files can also be uploaded. It will automatically fall back to that topic, use those data sources, and try to answer the question. If not, it will say, “Unfortunately I can’t help you about that.”
The current supported data sources are, as mentioned, public websites, SharePoint. Open AI deployments from Azure can be integrated. Bing Custom Search can be used, and also supplying custom data to these nodes if data is read from a different data source than those listed. Michael has some examples of this as well.
In the first demo, an integration with a public website was shown. A question was asked: “Tell me about Rockhop.” The Rockhop website was set as the public website, so it went and grabbed information from it and gave a summary. It also provided a link to where the information was found.
Another question was asked: “How can I contact Rockhop?” It returned the phone number, email, how fast the response is, and other places to connect. This was scraping the page and summarizing the information, with citation.
Another simple yes/no question was asked: “Does Rockhop do Power Platform work well?” The answer was yes, with citation.
Then a fun question: “What does Todd Golden like to do in his free time?” The answer included that he enjoys spending time with his family on the water and playing golf.
Next, the custom data piece was covered. There is a specific structure for the custom data field: an array of objects with two properties — content and content location — so that the source can be cited. There are multiple use cases for this. Static information can be placed in the custom data field. In the example, webinar details were added. This allows the generation of responses based on the provided information. SharePoint lists can also be used, with flows that pull data and return it in the correct format.
A question was asked: “When is the webinar?” The system returned: February 2nd at 10 a.m. Central. The variable containing this information was set up in Co-pilot Studio and fed into Generative Answers.
Next, SharePoint was demonstrated. A document library with invoices and rental agreements was set up. The question: “What was the subtotal for invoice 1726?” returned the exact subtotal. The invoice could be pulled up to verify.
Another question: “Let me contact the salesperson” returned the salesperson name for invoice 1726.
A deeper question: “Was a 25 pack of golf included on this invoice?” confirmed yes, with rate, quantity, and total. Handwriting in the invoice number was successfully read and parsed.
Questions about rental agreements followed. “What’s the rent price for 7890 Hill Street?” returned no result. “What’s the security deposit for 7890?” returned a specific number, located within the document.
Finally: “Who is currently living at 2345 Maple Street?” returned “Graham Barnes.” This required parsing a lot of text in the contract.
This SharePoint demo showed Generative Answers doing a good job of sifting through large documents and returning accurate results. This capability is powerful for quickly surfacing data from many documents within SharePoint sites.
