5.1.3 · Lesson 3 of 14
Score Grounding, Usefulness, and Safety
Score GenAI outputs for faithfulness/groundedness, answer relevance, and safety dimensions using LLM-as-a-Judge rubrics.
WHAT
Score completions for grounding, usefulness, and safety beyond lexical overlap.
HOW
- Faithfulness / Groundedness: Are claims derived from retrieved context?
- Answer Relevance: Does the completion address the user prompt?
- Context Precision & Recall: Did retrieval pull needed facts without noise?
- Use a stronger FM as judge against defined rubrics.
EXAM TRAP
RAG evaluation is not only factual correctness of final text — measure retrieval quality and groundedness separately.
Key points
- Faithfulness / Groundedness
- Answer Relevance
- Context Precision & Recall
- Toxicity / safety scores
- LLM-as-a-Judge rubrics
Exam focus
RAG evaluation needs three stages: context relevance, groundedness/faithfulness, and answer relevance — not only final-text correctness.