I think that "next level" would essentially be a RAG referential information system that gets data via search engines or databases. Maybe we will have the "google search" equivalent completely intended for LLM clients where all data is stored, searched, and returned via vector embeddings, but it could tap on exabytes of information.