Interview Guides

You are building a retrieval augmented generation system over a very large document corpus. Users ask natural language questions and expect grounded answers with relevant citations from enterprise content.
How would you architect a RAG system that handles millions of documents with low latency?