RAG Evaluation

The knowledge-assistant team from Chapter 9 has a candidate release: a new reranker and a rewritten permission-filter compiler. The regression dashboard looks excellent. On the judged questions, nDCG@5 rises from 0.877 to 0.973, faithfulness from 0.72 to 1.00, and a 60-question accuracy check shows the candidate five points ahead of production. The release manager asks the only question that matters: ship it?