Amey-Thakur RAG freshness
A retrieval index is a snapshot that starts drifting immediately. The failure is quiet and serious: an answer confidently cites a policy that was superseded last month, with a citation that makes it look verified.
Method
- State the staleness budget per source. Minutes for operational data, days for documentation. It determines the whole update strategy (see search-indexing-pipeline).
- Update incrementally on change events. Reindexing everything on a schedule is expensive and still leaves a window; capturing changes as they happen is both cheaper and fresher.
- Propagate deletions immediately. A removed document that stays retrievable is worse than a stale one, especially where the removal was for accuracy or permission reasons.
- Re-chunk and re-embed on substantive change. A minor edit may not warrant it; a rewrite does, and treating all edits alike either wastes work or leaves stale vectors (see chunking-strategies).
- Carry the source timestamp into the chunk. The model can then qualify its answer, and users can judge for themselves (see citation-grounding).
- Detect drift by sampling. Periodically compare indexed content against the source to catch pipeline failures that produce silent staleness.
- Alert on indexing lag. A stalled pipeline looks identical to a working one from the answer side until someone notices a wrong answer.
Boundaries
Freshness costs indexing work proportional to change rate. Some sources provide no change notification and require polling. An index can be perfectly fresh and still retrieve poorly, which is a separate problem (see rag-evaluation).