Citations and groundedness

Groundedness means the answer sticks to retrieved sources. Citations show which chunks support which claims so a human can verify.

RAG without citations is just a smoother hallucination surface. Products that matter (policy, medical adjacent, legal-ish) need checkable trails—not vibes.

1. Retrieve chunks
2. Require the model to quote or point to chunk IDs
3. Refuse or hedge when chunks don’t support the claim
4. Let users open the source

Measure groundedness with human review and automated claim-check experiments—not demos alone.

A student essay with footnotes beats a confident paragraph with no trail.

Ask a RAG-style assistant for a claim *and* the source snippet. If it won’t show a snippet, treat it as unverified.

⚠️ Myth: “According to the documents” guarantees truth.

✓ Reality: Documents can be wrong, outdated, or mismatched to the question.

⚠️ Myth: More citations always mean better answers.

✓ Reality: Citation spam can hide weak support.

• Course 09 RAG hub; hallucinations chapter
• NIST AI RMF (transparency themes): https://www.nist.gov/itl/ai-risk-management-framework ↗
• Provider groundedness eval guides (cite the one you use)