Skip to main content

GriMoire overview

GriMoire is an AI assistant that applies Reflexive UI Awareness inside SharePoint. It uses the user's delegated Microsoft 365 context, renders typed UI blocks, and can move from search to recap to action without switching tools.

The pattern behind that behavior is the Hybrid Interaction Engine (HIE). The HIE keeps the assistant grounded in the UI it creates. Instead of treating the chat transcript and the visible interface as separate worlds, GriMoire keeps them in the same loop — a pattern described as reflexive UI awareness.

note

In GriMoire, UI state is part of the interaction model, not just the output surface. The model receives visual context and interaction events from the blocks it creates.

What GriMoire is built to do​

  • Search Microsoft 365 content with an intent-based search pipeline that routes by intent, expands queries through LLM-based planning, and fuses results from multiple sources.
  • Render results as UI blocks that users can inspect, click, and act on.
  • Use Agent 365 and MCP-backed capabilities to bridge from information to actions such as mail, meetings, Teams operations, file operations, and more.
  • Stay in the user's delegated Microsoft 365 context for Microsoft Graph access (including Copilot endpoints) while keeping model access behind a backend proxy.

What you can try once the setup is complete​

  • Search SharePoint results, generate a recap, then say "send the recap by mail".
  • Say "find the budget docs and send them to the team" — GriMoire plans and executes the full chain automatically as a compound workflow.
  • Find a document and share it with another audience.
  • Summarize a Teams discussion and create a follow-up meeting.
  • Search an email thread, summarize it, and draft a reply.
  • Pull recent documents, summarize one, and send an update to a team.

The shape of the solution​

GriMoire has four layers:

  1. SharePoint web part frontend: the assistant UI, the block system, the expressive avatar, the HIE loop, and the delegated Microsoft Graph calls (including Copilot endpoints).
  2. Backend proxy: Azure Functions for model access, speech synthesis, MCP session execution, and Easy Auth-protected user data persistence.
  3. Intent-based search: intent routing, LLM query planning, multi-source retrieval (Copilot Search, Copilot Retrieval, SharePoint Search), and RRF fusion.
  4. Agent 365 and MCP tooling: SharePoint, OneDrive, SharePoint Lists, Outlook Mail, Outlook Calendar, Teams, Word, User Profile, and Copilot Search — exposed as governed tools through 8 MCP servers.

What GriMoire showcases​

GriMoire demonstrates how several Microsoft 365 and Azure AI technologies work together in a single assistant experience:

  • Intent-based search with result fusion — Copilot Search (semantic + lexical), Copilot Retrieval (RAG chunks), and SharePoint Search (KQL) run in parallel. An LLM-powered query expander rewrites the user's input into source-specific queries. Results are fused using Reciprocal Rank Fusion (RRF).
  • Copilot Chat API for document intelligence — Summarization, document Q&A, and conversational follow-ups use the Copilot Chat API (/beta/copilot/conversations).
  • LLM-based tool routing — The backend LLM decides which MCP tools to invoke based on the user's intent. The UI block rendering is deterministic — catalog hints and schema shape detection map tool results to the appropriate block type without LLM involvement.
  • Typed UI blocks with dynamic data — 16 block types (search results, file previews, user cards, compose forms, and more) are rendered based on the shape of the returned data.
  • Agent 365 MCP integration — 8 MCP servers exposing approximately 100 Microsoft 365 tools through a governed backend boundary.
  • Compound workflows — Multi-step intents ("find docs and send by email") are planned and executed as chained operations.
Project status

GriMoire covers a wide surface area across search, Microsoft 365 actions, voice, and UI rendering. Some capabilities may behave differently depending on your tenant configuration, model availability, or Agent 365 enrollment status. Review and test in your environment before production use. Contributions and feedback are welcome.

Setup​

The install guide walks through the full setup: deploying the backend, building the web part, uploading it to SharePoint, approving permissions, and configuring the Agent 365 tooling. Some steps require tenant admin access — the guide is explicit about what is needed at each stage.

Where to go next​

  • Start with Prerequisites if you want to confirm tenant and admin readiness.
  • Jump to Install if you are ready to set up the full stack.
  • Read Architecture and HIE if you want the conceptual model first.
  • See Avatar if you want to understand the expressive avatar system.