RAG Faithfulness

Detect hallucinations in retrieval-augmented generation. When a test case has a context field (the retrieved source documents), the judge runs an additional faithfulness check.

What it checks

Is every factual claim in the agent's response directly supported by the source context? No fabrication, no hallucination.

• faithful — all claims grounded in context
• unfaithful — contains unsupported claims (hallucination)
• partial — mostly grounded with minor additions

Stored as a separate AI score with a rag: prefix. Optional — only runs when context is present.