Llamar.ai: A deep dive into the (in)feasibility of RAG with LLMs