A citation can be real, clickable, and wrong.
Suppose an answer says your cancellation window is 30 days and cites the current supplier contract. The link opens. The document is genuine. The cited clause says only that either party may cancel in writing. The number came from somewhere else. The interface has provenance. The claim has no support.
LLM grounding is the support relationship between a generated claim and the evidence the model was given for that answer. A claim is grounded when one passage, or a stated combination of passages read together, entails what the sentence says, and when that evidence was present in the context at generation time. Plausibility does not pass. A similar passage elsewhere in the corpus does not pass. Training data does not pass. A tidy footnote does not pass either.
Grounding vs RAG
The words are used loosely, and the vendors do not agree with each other. Google’s Search guide describes retrieval-augmented generation as a technique “also known as grounding.” Google Cloud draws the relation the other way: RAG is a commonly used grounding technique. Both usages are current, and each is consistent inside its own product.
For operating a retrieval system, this site keeps the jobs apart:
| Job | Question it answers | Typical failure |
|---|---|---|
| RAG retrieval | Did the system find and assemble the relevant passages before generation? | The governing clause never entered the context |
| Grounding | Does each factual claim follow from the passages that entered the context? | The model adds a number the context does not contain, or joins two facts wrongly |
| Citation | Can a reader open the claimed support? | The link resolves, but the passage does not entail the sentence beside it |
RAG can retrieve excellent context and still produce an ungrounded answer. The model can skip the decisive sentence, import a familiar number from training, or attach the wrong source ID. Grounding tests that last mile. What is RAG? defines the retrieval architecture. This page defines the check that runs after it.
The inverse case matters too. A claim can be faithfully supported by an obsolete policy. It is grounded in the supplied source and still wrong for today’s decision. That is a freshness failure, not a grounding failure, and What is GRAIN? owns it.
Two failures that look alike
The book behind this page names two. Ungrounded generation is a claim with no support anywhere in the assembled context. The model produced a plausible continuation from its training distribution, and nothing retrieved can vouch for it. Miscited generation is subtler. The model cites a real passage from the context, but the passage does not say what the claim says. The syntax is correct, retrieval did its job, and the association is false.
Both are caught by the same mechanism: compare the output against the assembled context and ask, claim by claim, whether each follows from what was given. The check has to be structural. A prompt line that says “only use the provided context” is an instruction, not a guarantee.
The support test
Run it at claim level, not paragraph level.
- Split the answer into factual claims. “The policy changed on 1 June and now allows 30 days” contains at least two.
- Record the exact passages present in the generation-time context, with stable source and chunk identifiers.
- For each claim, name the passage or passages that support it.
- Ask whether the evidence entails the whole claim. Shared vocabulary and semantic similarity are weaker tests, and they pass sentences they should not.
- Verify that the visible citation opens those same supporting passages.
- Route what fails: drop a peripheral detail, answer only the supported part, retrieve again, refuse, or escalate to a person.
The order matters because of a common repair that changes nothing. A team finds an unsupported sentence, searches the corpus afterwards, finds a paragraph that could support it, and adds the citation. The display now looks defensible. The original generation was still ungrounded, because the evidence arrived after the answer. A source found later cannot ground a sentence written earlier.
Where GRAIN puts the work
In The RAG Engineer, GRAIN means Gather corpora, Rank and rerank, Assemble context, Inspect failures, Navigate freshness.
The grounding check belongs to Assemble, as integral to that stage as building the context itself. Source identifiers and passage boundaries travel into the context, and the check asks whether the resulting claims follow from it. Every flagged answer then becomes a record for Inspect, which keeps the failure trace and measures how often the check fires. Navigate keeps the supporting source current.
That division stops one overloaded word from hiding three different defects:
- the right evidence was never retrieved;
- the evidence was retrieved and did not support the generated claim;
- the evidence supported the claim when indexed and has since been superseded.
Each has a different owner and a different repair.
A citation is an interface, not a verdict
Clickable source IDs are valuable. They let a reader open the exact passage, and they let software confirm that an identifier existed in the assembly trace. That is referential validity: the pointer is real.
Grounding requires the second check. Does the pointed passage support the sentence beside it? A model can cite clause 12.3 accurately and misstate clause 12.3 in the same breath.
Automated grounding judges can triage this work. Their scores measure a particular judge, prompt, claim set, and threshold; they are not truth meters. Build labelled examples from your own corpus, include partial-support cases, and keep a human route for decisions where a polished error costs real money. Evaluate the reasoning, not the fluency covers the wider habit of mistaking finish for evidence.
Grounding sits between retrieval and trust. Every factual sentence must be able to reopen the evidence that was on the table when it was written.
Cite this:What is LLM grounding? The sentence has to open a folder.Len P. van der Hof. https://lenvanderhof.com/en/blog/what-is-llm-grounding/ ·