Current is a behavior you test
Define freshness as an observable behavior: after an approved change, a permitted user receives the new answer, the old answer is no longer retrieved for current questions, and the source link points to the right revision.
Give every source an owner and a status
Record source ID, owner, approval state, effective period, review date, audience, last successful processing time, and deletion status. Keep an untouched original and a traceable derived representation.
The owner decides whether a change is authoritative. The ingestion operator confirms that the change was processed. Those are different responsibilities and should not be hidden behind one green sync badge.
Test replacement and removal separately
A replacement test asks whether a new rule wins when the old rule is still nearby. A removal test asks whether deleted or revoked material stops appearing in fresh retrieval and existing conversations according to the system's documented behavior.
Microsoft's external-content documentation makes the item lifecycle explicit. Your exact API and indexer behavior may differ, so run the test against the configured product instead of borrowing a vendor promise.
Handle stale and conflicting material honestly
Use effective dates, source authority, scope, and review status as retrieval signals. If two documents are both current for different regions, preserve that distinction. If no owner has resolved the conflict, the answer should say so and route the question.
A model should not be asked to infer policy hierarchy from wording alone.
Make freshness part of release management
Keep a small regression set of questions for important policies and workflows. Run it after connector, parser, index, permission, and model changes. Record source IDs, answer versions, retrieval evidence, and unresolved failures.
A failed freshness test is a release signal, not a reason to hide the old answer with a stronger prompt.
Design the freshness pipeline
Separate source change detection, extraction, indexing, cache invalidation, and answer evaluation. Record a processing ID and timestamp at each stage. When a source is replaced, test that the new version is retrievable and that the old version no longer wins for a current question. When a source is deleted or access is revoked, test direct retrieval, related questions, and an existing conversation.
Use different service targets for low-risk reference material and high-risk ownership or access data. A daily refresh may be fine for a handbook and unacceptable for a revoked employee account.
Resolve conflicts in the open
Two current documents can disagree because they have different audiences or approval states. Keep scope, effective dates, and authority visible. Route unresolved conflicts to the source owner and prevent the assistant from silently choosing the most recent file.
Review freshness failures as operational incidents. A green connector status is not proof that the answer layer is current, so report stale-source examples and the time needed to repair them.
The LeanOrchestr Business Brain connects approved company sources, preserves access boundaries, and turns useful context into a maintained operating layer. Start with one question, one accountable owner, and an evaluation set you can rerun.
Sources and verification
Primary product, architecture, risk, and security sources checked on 10 September 2026. Product behavior can change, so verify current plan and admin documentation before implementation.