Why it matters
Stale knowledge is outdated but still retrievable. It causes wrong answers without the drama of hallucination. Customers blame the company, not the timestamp field you forgot to set.
How it works
Detect staleness via review SLAs, monitoring diffs, and failed tests. Quarantine or downgrade stale chunks in retrieval ranking. Force refresh workflows.
Example
Nintendo retires a stale 30-day refund FAQ still indexed from a migration. Until removal, metadata marks it non-authoritative so ranking deprioritizes it against current Refund Policy chunks.
Common mistakes
- 1Leaving legacy collections indexed for sentimental reasons
- 2Stale tags with no retrieval effect
- 3No comms to support when stale content removed
Your source changed. Your AI should know.
Track source freshness, detect changes, and review impact before answers drift.
Explore Source Maintenance