A Tale of Trust and Accuracy: Base vs. Instruct LLMs in RAG Systems