Groundedness Check Prompt
Verify a RAG answer is actually from its retrieved context — every claim traced to a retrieved passage, and any answer that outran its sources flagged.
Overview
RAG's whole promise is grounded answers, but a model will happily blend retrieved context with its own training and present the mix as sourced. This prompt checks groundedness specifically: it maps each part of the answer to the retrieved passage that supports it, flags claims with no supporting passage, and catches the case the system should have handled — answering confidently when the retrieved context didn't actually contain the answer.
Why This Works
- Attributing each claim to a passage is the only real test of grounding
- Catching confident-answer-without-context targets RAG's signature failure
- Distinguishing 'wrong' from 'unsupported-but-right' isolates the retrieval problem
Best for
- RAG systems and knowledge-grounded agents
- Domains where an unsourced answer is dangerous
- Evaluating retrieval quality and answer faithfulness together
Not for
- Checking output against an arbitrary source — use the Hallucination Detection Prompt
- Non-RAG generation with no retrieved context
Use cases
- Verifying RAG answers are sourced, not invented
- Catching answers that outran the retrieved context
- Evaluating retrieval-augmented agents for grounding