Embeddings vs LLMs: Why Does RAG Need Both?

I separated retrieval and generation in a small RAG experiment to see what embeddings and LLMs actually do, and why both matter.