What is grounding?
A chatbot can retrieve the right document and still write an answer that document never said. This is the check that catches that gap.
Grounding
Grounded generation
Grounding is the link between a claim in a generated answer and the source it came from. A reader can trace the claim back to its origin. A grounded answer is checkable; an ungrounded one is a guess dressed up as a fact.
Retrieval-augmented generation is the mechanism. Grounding is the property it is supposed to produce. A RAG system fetches documents and hands them to a model, but nothing forces the model to use only what it was handed. It can still blend in something from training data, or state a conclusion the source only implies.
That gap is what grounding closes. A grounded system links each sentence of an answer to the passage it came from, and drops or flags any sentence with no matching passage. Retrieval fetches the material. Grounding proves the answer stayed inside it.
The confusion between the two terms is common because most demos skip the proof step. Retrieval without a citation check is decoration. The documents sit in the prompt, the answer sounds informed, and nobody checked whether the model actually used them.
An answer without a source is a claim, not a fact
A model can produce a fluent, confident sentence whether or not the retrieved documents support it. Fluency and accuracy are separate things, and nothing about how a language model writes forces them to match. Grounding is the step that tests the match, not the wording.
On our ESG research platform, Greenlight, this is the whole product, not a feature bolted on. Each company scan covers more than 50 sustainability topics, and every topic carries 10 to 15 independent sources with live citations attached to the finding. A user can click through to the actual document behind any score.
That pipeline runs retrieval grounding on purpose, because the alternative in ESG scoring is trusting a company's own claims about itself. Sources get weighted by trust and freshness before the model ever writes a sentence, and an expert council reviews the scoring. Grounding was the design constraint, not an add-on after launch.
- 01Retrieval finds candidate documents; a separate check confirms the answer stayed inside them.
- 02A system can retrieve correctly and still generate an ungrounded sentence.
- 03Citations at the span level, not a source list at the bottom, are what make grounding checkable.
- QueryThe question a user asks.
- RetrieveFetch candidate source passages.
- GenerateModel drafts an answer from them.
- VerifyEach claim checked against a source.
- CiteUnsupported claims cut or flagged.
The verify step is what separates a grounded answer from a plausible-sounding one.
Common questions
01Is grounding the same thing as RAG?
No. RAG is the retrieval-and-generation mechanism; grounding is the property of an answer staying inside what was retrieved. A RAG system can still produce an ungrounded answer if nothing checks the output against the sources.
02How do you verify grounding in a production system?
With span-level citation checking. Each sentence gets matched against a retrieved passage, and any sentence with no match is cut or flagged before it reaches the user. A source list at the bottom of the response does not do this. It only shows what was fetched, not what the answer actually used.
03Does grounding eliminate hallucination?
It reduces it sharply but does not eliminate it. A model can still misread a source it correctly retrieved, so if you are evaluating a vendor's claim, ask how they catch that case. Grounding narrows the failure from 'invented from nowhere' to 'misread a real document', which is smaller and easier to catch.
04Do regulated industries require grounding, or is it optional?
Buyers in healthcare, finance and legal work increasingly treat grounding as a requirement, not a nice-to-have. An unsourced claim in those fields carries real liability, and a system with no citation trail is hard to defend in an audit.
