What are the pros and cons of (1) fine-tuning a base model on proprietary data, (2) using retrieval-augmented generation (RAG) and agentic workflows with a general-purpose model, and (3) combining both approaches? Links to resources discussing this would be awesome.