A sourced answer.
But was the source authorized?

The shared folder is not a single block

A company often has shared documents, restricted areas and files accessible to a single team. Copying all these files into a search index must not erase these distinctions. The relationship between document, source and viewing right must be kept.

The Microsoft guide to document-level access controls describes several filtering and authorization mechanisms in Azure AI Search. Their availability depends on the option used, and some are in preview. They illustrate an architectural need; they do not prove that your engine or connector already covers it.

A test with two roles

Build a fictional corpus with a procedure open to the whole team and a note reserved for a management role. Ask the same question from both identities. Compare the retrieved passages, the citations and the final answer. A title, an excerpt or the file name can already reveal information.

Then remove access to the note and run the test again. Also check the caches and the histories: a change in the original directory does not necessarily erase all the copies already created.

Plan for the document life cycle

• The original source and its version must remain identifiable.
• An update must replace the previous version or be clearly distinguished from it.
• A deletion must have a defined effect in the index and in the working copies.
• An expired document must be able to be excluded or clearly flagged.
• Questions without a sufficient source must not end up with an invented conclusion.

Do not ask the prompt to do all the checking

An instruction saying “do not reveal private documents” does not replace a check at the moment the passages are retrieved. The application must verify the rights before giving the information to the model, then control what it keeps and displays.

For a first project, Saphan can work on a limited corpus, explicit roles and contradictory scenarios. The size of the knowledge base comes after the ability to know who sees what.