RAG optimization — chunking and embedding quality are indexing-time problems; hybrid search, query rewriting, and reranking are retrieval-time problems; context ordering and CRAG are generation-time problems. Most teams only touch retrieval and miss the 60% of the wins available at the other two layers. Stale data — three defenses. Event-driven ingestion so you don't rely on nightly batch. Document-level versioning so re-ingestion atomically replaces old chunks. Temporal metadata on every chunk so retrieval can filter or score by recency. The hardest case is facts with known validity windows — that's where a temporal graph like Zep earns its complexity