How to evaluate an LLM-powered RAG application automatically.