The challenge
Years of proposals, technical reports and internal procedures were spread across SharePoint sites and file shares. Staff spent a lot of time hunting for past work, and new joiners relied on asking colleagues.
Leadership wanted to try AI but was wary of confident-sounding wrong answers, and of documents being shown to people who shouldn’t see them.
Our solution
A short discovery picked the most valuable document collections and gathered real questions staff ask. Documents were split into passages, embedded and indexed in Azure AI Search, with every passage carrying the access permissions of its source document.
Staff ask questions in a Next.js web app that is also available as a Microsoft Teams tab. Answers come from Azure OpenAI using only passages the user is allowed to see, always with citations, and the assistant says when it can’t find an answer rather than guessing.
- Retrieval that respects existing SharePoint permissions
- Every answer cites the source passage
- Evaluation set built from real staff questions
- Usage, feedback and cost dashboard by department
Architecture
Sources
- SharePoint
- File shares
Indexing
- Chunk + embed
- Azure AI Search
App
- Next.js + Teams tab
- Entra ID SSO
AI
- Azure OpenAI
- Evaluation & logs
Technology
- Azure OpenAI
- Azure AI Search
- Next.js
- TypeScript
- Microsoft Entra ID
- SharePoint
- Azure Functions
Outcomes
- Staff find past work and procedures without asking around
- Every answer can be checked against its source document
- Leadership can see usage, quality feedback and running costs in one place