Four-Stage Retrieval Design for News Monitoring with Query Limits and Provenance

The article proposes a four-stage retrieval architecture for news monitoring services: collection, chunking, indexing, and retrieval. It argues that stale or unverifiable answers should be treated as ingestion and retrieval-contract failures rather than prompt problems. Requests must specify a collection, access scope, bounded result count, and source context, with unbounded queries rejected early. The author recommends provider-neutral conformance tests, honoring Retry-After on rate limits, and deterministic chunk IDs that allow safe re-ingestion. Index changes are proposed to roll out through warmed generations and alias switches, enabling simple rollback.