Why retrieval instead of fine-tuning the model on this...

A strong answer focuses on the operational transfers: correctness, attribution, opt-out, adaptability, and parameter efficiency. The definitive argument for RAG over fine-tuning is the requirement for bounded deletion. Because unlearning requires the developer to enumerate exactly what to forget, it is not a viable strategy for compliance. Relying on the claim that ‘RAG is cheaper’ or ‘reduces hallucination’ is insufficient, as it ignores the architectural complexities of maintaining a production retrieval system.