RAG systems fail quietly when the knowledge base becomes unmanaged. Policies change, documents are superseded, permissions evolve, and old pages are archived. If your RAG pipeline treats content as static, users get stale answers, broken citations, or access-control incidents.
Lifecycle management is the discipline of treating every source as a product with ownership, change control and retirements.
Assign ownership and review cadences per source
Every source system and major collection should have:
- Named owner. Who is accountable for quality, permissions and updates.
- Review cadence. How often content and access rules are reviewed.
- Change signals. Webhooks, events, or scan patterns to detect updates.
This is part of connector hardening (see hardening connectors and content review workflows).
Model the lifecycle states explicitly
A simple lifecycle state model prevents ambiguity:
- Active. Used for retrieval and citations.
- Deprecated. Still retrievable, but should be down-ranked or flagged; often replaced by a newer policy or document.
- Archived. Not retrievable by default; retained for compliance or history.
- Deleted. Removed from retrieval and indexes, with tombstones to prevent reappearance.
These states can be represented as metadata fields and enforced in retrieval filters and ranking.
Handle supersession: the "new policy replaces old policy" case
Supersession is a common pattern in enterprise content. If a new document replaces an old one, do not simply delete the old content. Instead:
- Mark the old document as deprecated and link it to the replacement ID.
- Down-rank deprecated sources unless the question is explicitly historical.
- Prefer the newest policy when answering time-sensitive questions (see freshness evaluation).
Make deletes reliable end-to-end
Deletes are not just "remove from the vector DB". You need an end-to-end deletion workflow:
- Delete from the raw store, the chunk store and the vector index.
- Invalidate caches and derived indexes.
- Record tombstones so re-ingestion does not resurrect content.
Design deletes as a first-class pipeline (see deletion workflows and data retention and deletion).
Keep access controls in sync throughout the lifecycle
Lifecycle changes often correlate with permission changes. Ensure:
- ACLs are captured and stored as primary metadata.
- Retrieval filters enforce entitlements, not just ranking (see RAG permissions).
- Archived sources have stricter defaults to avoid accidental disclosure.
Measure freshness and lifecycle health
Lifecycle management becomes real when it is measured. Useful signals include:
- Freshness by domain. How old are the sources used in answers?
- Deprecated usage rate. How often do deprecated sources appear in top-k retrieval?
- Delete lag. How long does it take for deletes to disappear from retrieval?
Make these part of your observability and alerting (see freshness architecture and alerting and runbooks).
RAG trust is a content operations problem as much as a model problem. When you manage the lifecycle, the system stays accurate as the organisation changes.