How do you build a ground-truth dataset to evaluate if your RAG system is actually improving over time?

If the retriever returns zero relevant documents, how do you prevent the LLM from making up an answer?

If your retriever finds 50 relevant "top" chunks but your LLM context window only fits 10, how do you decide which ones to keep?

Walk me through the lifecycle of a user query, from the moment they hit "Enter" to the final response generation.