I wrote this after shipping multiple RAG systems that broke in production. This post focuses on architecture choices, failure modes, and tradeoffs. No prompt tricks. No hype. Feedback from people running RAG in production is welcome.