What I Learned from Building Google Search and How It Inspires RAG Development

Very interesting comparison of RAG to traditional search development. I'm curious how large an evaluation dataset is usually necessary to compare results from different experiments of Google search?