Building a RAG assistant with Microsoft Graph and .NET
Microsoft Graph supplies calendar, mail and document data that can be indexed with Azure AI Search and passed to a language model to produce sourced answers. The article walks through a .NET retrieval-augmented generation flow with code for loading context, chunking documents and calling the model. It stresses least-privilege scopes and permission-aware retrieval, which matter for any enterprise assistant.
Key points
- Keep Graph permissions as narrow as possible, such as Mail.Read, Calendars.Read and Files.Read.
- Map mail and calendar items into one shared context model with source, title, text and URL.
- Split large documents into overlapping chunks, embed them and store them in an Azure AI Search index.
- Send only a few relevant chunks to the model and keep source URLs for citations.
- Re-check access rules at retrieval time so the index never returns content the user cannot read.
- Require explicit user confirmation before the assistant sends mail or changes calendar events.
Generative AI becomes useful in business apps only when it knows what the user is working on. This article shows how a .NET application can combine Microsoft Graph data from Outlook, calendar, OneDrive and SharePoint with Azure AI Search and a language model to produce grounded, source-referenced answers. It is a retrieval-augmented generation pattern in which Graph is the source of permitted context rather than the AI service itself.
The author covers the practical building blocks: narrow Graph scopes, a small shared context model for mail and calendar items, text extraction and overlapping chunks for documents, and a hybrid vector and keyword query against the index. On the model side, Microsoft.Extensions.AI with Azure OpenAI and a managed identity avoids storing API keys. The piece also argues for permission-aware retrieval, data minimization, treating document content as untrusted input, and human confirmation before the assistant sends mail or changes events.
Real-world use
- An internal assistant that prepares a customer meeting by summarizing the calendar, recent mail and project documents.
- An enterprise question-answering app over a large SharePoint library that cites the source documents.
- A Microsoft 365 integration that drafts Outlook replies while leaving the final send action to the user.
Why it matters
Any .NET team connecting Microsoft 365 data to a language model needs to design scopes, data minimization and permission-aware search from the start. These decisions determine whether the prototype can safely become an enterprise assistant.