Incremental RAG indexing that keeps the previous index intact

Re-indexing a live retrieval collection is hard because documents, parsers, embedding services, and vector storage can fail at different points. Updating chunks one at a time can expose users to half-updated or inconsistent data, undermining trust in answers. A safer approach separates preparing a new document version from publishing it, using a small indexing record per source with stable keys, content hashes, and lifecycle states. The Project Pavilion example on .NET 9 and Qdrant overwrites a single point per record, so retrieval never sees a mix of old and new vectors. Publication happens through one durable active-version switch, while stale vectors are cleaned up later and the last known-good index serves as a labelled fallback during outages.