What it is: You assess the final response generated by the entire RAG pipeline.
Why it matters: The responses can vary (non-deterministic), so you need to check if the output aligns with a "ground truth" or the ideal response.