@Heymaxi01 LLMs don't take vector embeddings as input, but internally it still will convert natural language -> tokens -> embeddings.
In RAG systems, we can perform text-based searches (keyword search) but it may miss relevant information as compared to embedding-based retrieval.