�ݺ�ߣ

Editor's Notes

• #4: Here is an overview of the Copilot architecture that your just saw. A few things I want to call out. In this example, we are calling the Microsoft Graph for grounding data, but we are also working on including Web data and even plug-ins, which there will be a deep dive on later. I also really want to emphasize this because it seems to be a big point of confusion with our customers: all this occurs within the Microsoft 365 Trust Boundary. Any content that’s processed is limited to what users have access to, and generated content becomes content in Microsoft 365, and thus has the same security, privacy, and trust promises that any data in Microsoft 365 has. Although the buzz around GPT and LLMs is new, the components that power Copilot are foundational platform pieces that we’ve been working for years on. As I mentioned earlier, the Copilot Orchestration engine also has additional RAI checks, so the central spinning image you see here is actually doing a lot of heavy lifting. Now that you have the overall picture, let’s deep dive into the area I’ve been working for the last 7 years: the Microsoft Graph and Semantic Search. ---- Standalone talking points for this slide: Let’s take a look at a Copilot agent architecture. For example, in the bizchat flow, a user might ask a question. The user prompt is sent to the M365 Copilot Platform. The orchestrator understands what the user is trying to do and can evaluate what grounding data it needs to retrieve from the graph or web, and what plugins it might need to run to accomplish the task. Copilot sends the modified prompt to the LLM which generates a response.