Grounding
Grounding is the practice of tying a model’s output to specific source material supplied with the request, so that each claim in the answer traces back to provided text rather than to whatever the model absorbed during training.
A language model asked a question with no material attached answers from its parameters. Whatever it produces may be correct, but nothing in the system can say where it came from or check it. Grounding changes the setup: relevant text is placed in the prompt and the model is instructed to answer from it, which makes the answer’s basis inspectable.
The word describes a relationship, not a mechanism. Retrieval is the usual way to supply the material, but pasting a document in by hand grounds an answer just as much. What makes an answer grounded is that the text it depends on was provided and can be pointed at.
Grounding is a matter of degree and it is not self-verifying. Supplying source material makes an answer capable of being grounded; whether the model actually used it, and used it correctly, is a separate question requiring separate evidence.
The term is borrowed from cognitive science and linguistics, where it concerns how symbols acquire meaning by reference to something outside the symbol system. The retrieval sense keeps the shape of that idea and narrows it to documents.
In practice
Grounding shows up as three separate things in a retrieval-augmented system, and they are often discussed as one.
Supplying the material. Retrieved passages are placed in the prompt, usually with their source identified. What is present here is the ceiling on what can be grounded: a claim whose support was not retrieved cannot be grounded, whatever the instructions say.
Instructing the model. The prompt states that the answer should come from the supplied text, commonly with a provision for saying nothing is available. Such instructions shift behaviour rather than constraining it — the model retains access to its parameters and no wording removes that.
Attribution. The answer carries markers pointing to which supplied passage supports which statement. Attribution is what makes grounding checkable by a reader, and it is also the part most easily faked: a citation marker is text the model generated, and its presence is not evidence that the cited passage says what the sentence claims.
Where the material and the model’s parameters disagree, the outcome is not fixed. A model may follow the supplied text, follow its training, or blend the two, and which happens varies with the model, the prompt, and how confidently the conflicting claim was represented in training. Outdated documents in a corpus surface this directly.
Commonly confused with
Retrieval-augmented generation. RAG is an architecture; grounding is the property the architecture is built to obtain. A system can implement RAG and produce poorly grounded answers, if the retrieved passages are irrelevant or the model disregards them. See semantic search.
Faithfulness. The measurement, not the practice. Faithfulness asks how much of a given answer is actually supported by the supplied material. Grounding is what the system attempts; faithfulness is what an evaluation finds.
Accuracy. A grounded answer is faithful to its sources, which is not the same as being true. If a supplied document is wrong, an answer that reproduces it is well grounded and false. The two properties are independent, and conflating them is the most consequential error the term invites.
Citation. A presentation feature. Citations display the claimed basis of an answer; grounding is the relationship they claim to show. A cited answer is not necessarily a grounded one.
Hallucination. The failure the practice is meant to reduce, defined by output not supported by anything available. Grounding narrows the opportunity for it without eliminating it. See hallucination.
Usage notes
“Grounded” is used as a product claim with no defined threshold. A system described as grounded has usually been built to supply retrieved material, which is a statement about its architecture rather than about its output. No specific rate of supported claims is implied.
In multimodal and robotics writing the word means something else. There, grounding refers to connecting language to perception or to physical action — associating the word for an object with the object. That usage is older within machine learning and unrelated to document retrieval.
“Grounding data” and “grounding source” name the supplied material in some vendor documentation, where the noun refers to the corpus rather than to the relationship. The same documentation often uses the verb in the sense given here, on the same page.
Grounding does not imply the material was read in full. Long contexts are used unevenly, so a passage present in the prompt is not necessarily a passage the answer accounts for. See context window.