Retrieval drift appears when previously reliable questions start returning weaker evidence, stale snippets, denied sources, or unsupported answers. It is usually gradual and easy to miss without operating metrics.

Track known-answer probes, citation precision, denied-source behavior, freshness misses, and reviewer corrections by source class. Compare those signals by release version and ingestion window.

Remediation should name the failing layer: source update, parser, chunking, permissions, ranking, prompt, or reviewer policy. That separation keeps teams from treating every retrieval problem as a model problem.