Explain how vector embeddings and similarity search enable semantic document retrieval in a RAG system.