A private AI knowledge base with permission-aware search
A private knowledge system is useful only when answers cite evidence, revoked documents disappear and every user sees only material they are authorised to read.
Start with governed sources
List repositories, owners, versions, sensitivity and freshness rules. Pilot a narrow subject with approved content rather than copying an entire file server into one index.
Store source owner, version, location, validity and visibility beside each indexed fragment, not only inside the original repository.
Filter before retrieval
Enforce source permissions before any fragment reaches the model, and synchronise groups, revocations and deletions from the system of record. Instructions to the model are not access control.
Synchronise grants and revocations before retrieval; asking the language model to respect a label is not technical access enforcement.
Return evidence
Attach source title, version, stable location, supporting passage and update date. Refuse to invent an answer when retrieval has no adequate evidence.
Show the supporting passage and current source version, and decline when the available material does not establish an answer.
Evaluate two systems
Measure whether retrieval found the permitted current source separately from whether generation used it accurately. Include old revisions, scans, tables, absent answers and hostile instructions in documents.
Test retrieval and generation separately with old revisions, scans, tables, missing evidence and hostile instructions embedded in content.
Need an estimate for your project?
Tell us about the project. We will break it into stages and explain the budget drivers.