Grounding and citations

Lesson 3 of 5 in RAG for Agents: From One-Shot Retrieval to Agentic Search.

Retrieval was supposed to fix hallucination — and it helps, but it does not finish the job. Put the right passages in context and the model can still ignore them, blend them with its training-time beliefs, or over-claim beyond what they say. Grounding is the property you actually want: every claim in the answer is supported by a retrieved source. Citations are how you make that checkable — each claim points at the passage that backs it, so a human (or another system) can trace the answer to evidence.

For agents the stakes are higher than for chatbots. A chatbot’s ungrounded sentence misleads a reader; an agent acting on an ungrounded ‘fact’ — refunding against a policy clause that does not exist — turns the hallucination into a transaction.

Verification comes in two tiers, and knowing which is which keeps you honest.

Tier 1 — mechanical quote checking. If the system quotes its sources, you can verify each quoted span actually appears in the cited document with plain string (or fuzzy) matching. No model needed, runs in microseconds, catches outright fabrication. This is why well-designed RAG systems make the model quote verbatim — it converts an unfalsifiable claim into a checkable one.

Tier 2 — graded groundedness. The quote can be real and the claim still wrong: quoted out of context, stale, or stretched past what it says. Judging support — does this passage entail this claim, partially support it, or contradict it? — takes reading comprehension, so it is scored by an LLM-as-judge, claim by claim, in your evals. The evals domain covers how to calibrate that judge; here, just internalize that ‘groundedness’ is graded, not binary.

Interactive sorting exercise: Mechanical check or judge-graded? Sort each verification step.

Key terms: grounding, citation, groundedness, hallucination, LLM-as-judge, provenance

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.