Long Context RAG Performance of Large Language Models