Ground it in your data
"Grounding" means giving Copilot real source material to work from. A grounded answer is built from your documents, email, meetings or code. An ungrounded one relies on the model's general knowledge, which is where generic and made-up answers come from.
How Microsoft 365 Copilot finds your content
Microsoft 365 Copilot retrieves relevant content from Microsoft Graph (your files, mail, chats, calendar and meetings) before it generates a response. It only retrieves content the signed-in user already has permission to see. Microsoft's Copilot overview explains this flow in detail.
Retrieval is good, but it isn't psychic. If several documents could match, Copilot may pick the wrong one. Naming the source removes the guesswork.
Point to a specific source
- Name it in the prompt: "Using
Q3 Budget.xlsx..." or "Based on the email thread with the vendor about renewal..." - Reference it directly: in Microsoft 365 Copilot Chat you can attach or reference a specific file, email or meeting so it's used as the source.
- Work inside the document: Copilot in Word, Excel or PowerPoint already uses the open file as its primary context.
- Choose work vs. web: if your organization offers both, use work-grounded chat for internal questions and web-grounded chat for public information.
Grounding GitHub Copilot
In the editor, context comes from the code around your cursor, your open files and anything you explicitly reference in chat. To keep suggestions on track:
- Reference the specific files or symbols you're asking about in Copilot Chat.
- Add repository custom instructions so Copilot knows your stack, conventions and build commands.
Grounding your own agents
If you build agents in Copilot Studio, you choose the knowledge sources yourself, such as SharePoint sites, public websites or uploaded files. Narrow, well-maintained sources produce more reliable agents than "everything we have."