Internal Copilot on Private Knowledge
A copilot over internal knowledge that respects the permissions people already have, so it can never surface a document someone was not allowed to open. Answers cite their sources, staleness is visible, and the questions it cannot answer well are the ones it hands to a person.
- To first release
- 2 weeks
- Permissions
- Per-user
- Claude
- pgvector
- Next.js
- OpenTelemetry
The problem
Institutional knowledge is spread across wikis, drives, and chat history. A naive index over all of it becomes a permissions incident on day one, because retrieval ignores who is allowed to see what.
How we would sequence it
- 01Mirror source-system permissions into the retrieval layer itself
- 02Filter candidates by the asking user before ranking, never after
- 03Require citations so every answer is traceable to a document
- 04Pilot with one department before any org-wide rollout
- 05Log unanswered questions as a content gap backlog
This solution describes our method for a class of problem. It is not a record of a completed client engagement, and the figures above are scope estimates rather than measured results.
Have this problem?
Scope this with us