Retrieval-augmented generation works well in controlled demos, but often degrades in production. As data grows, changes frequently, or spans multiple sources, static retrieval pipelines introduce latency, inconsistency, and subtle correctness issues that are hard to diagnose. This session examines why first-generation RAG patterns break down in real systems. We’ll look at common failure modes stale context, over-retrieval, poor relevance, and hidden coupling between data and prompts and how teams are moving toward more context-aware designs. The focus is on treating retrieval as a decision, not a default: dynamically scoped access, owned and versioned context, and retrieval strategies that adapt to intent, cost, and risk.