Stop Overpaying for RAG: How We Cut Azure OpenAI Costs by 40% with One Architecture Tweak