RAG Development

SharePoint AI Search

SharePoint holds the documents and nobody can find them. The technical difficulty is not search quality — it is that SharePoint's permissions are layered and inconsistent, and an index that flattens them becomes a data leak.

Discuss your project

SharePoint permissions are the hard part

Access in SharePoint can be set at the site, the library, the folder and the individual item, with inheritance that may be broken at any level. Add sharing links — some anyone-with-the-link, some scoped to named people — and guest access from outside your tenant, and the effective permission on a given file is a computed result rather than a stored fact.

Any assistant over SharePoint must respect that computed result. We read effective permissions through the Microsoft Graph API, carry them into the index as metadata, and filter at retrieval against the identity of the person asking. A document the user could not open in SharePoint is never retrieved, so it cannot be quoted or summarised.

Sensitivity labels matter too. Where your organisation uses Microsoft Purview labels, documents carrying a restricted label can be excluded from indexing entirely rather than merely filtered, which is often what a compliance team will want. That is a policy decision we would ask you to make explicitly at scoping.

The practical reality is that many tenants have permissions nobody fully understands, accumulated over years. Indexing tends to reveal that, and we would rather surface it in a report during the build than have it surface as an answer.

What it does

Answers from documents

Word, Excel, PowerPoint and PDF content read properly — including text in slides and in spreadsheet cells, which keyword search handles badly.

Citations to the file and section

With a link that opens the document in SharePoint, so a user verifies rather than trusts.

Across sites and OneDrive

Within what the user may see. Covers the case where the right document is in a site the person did not know existed.

Says when it does not know

Rather than assembling a plausible answer from loosely related documents, which is the specific failure that destroys trust in these systems.

In Teams, where people are

A Teams bot or tab, since that is where your users already work. A separate portal gets opened once.

Metadata-aware

Site, library, document type and your own columns kept as metadata, so a question about a particular area can be constrained to it.

Common questions

Tell us what you need

Describe the problem and we will come back with scope, timeline, and a rate. No obligation to proceed.

Get in touch