What kind of abstractions around llm-io would you prefer?