Glossary
Grounding
Grounding ties an output claim to information that can be inspected outside the claim itself.
Context engineering
Grounding is the relationship between an assertion and material offered as its support. In question answering, a grounded response is constrained by identifiable sources or observations rather than presented as free-standing model knowledge. The strength of grounding depends on whether the cited material actually entails the claim, whether it applies to the correct entity and time, and whether its own status is understood. A citation marker alone does not establish any of these properties.
Practical grounding keeps the relevant passage, its source identity, and its locator close to each substantive claim. Reviewers should be able to distinguish quoted source content from interpretation and inference. Where two sources disagree, grounding requires showing the conflict rather than hiding it behind a single selected sentence. For numerical or legal conclusions, the calculation or interpretive step may need its own explanation even when the underlying inputs are cited.
Grounding matters because a decision maker may need to challenge a value, not merely read a convincing answer. Sentence-level citation is one visible method of connecting statements to evidence spans; retained source copies help those links survive changes to external pages. Faithfulness asks whether a generated answer stays within its supplied evidence. Grounding does not settle source reliability or resolve contradictions by itself, so it belongs within a broader evidence practice.
Example
A statement that Halden Therapeutics AG (fictional) held a licence on 3 May points to the licence's effective-period passage and the precise source locator.
What counts as grounded
A statement is grounded if a reader can inspect something outside the statement that supports it. For generated text the standard is made precise by Rashkin and colleagues, who define attribution to identified sources: a statement is attributable if a human, given the statement and an identified source, would judge that the source supports it. The definition separates grounding from truth. A statement can be attributable to a source that is wrong, and a true statement can be ungrounded because no source was cited. For organisations that must defend their outputs, grounding is the property that can be checked, and truth is the property that is ultimately wanted; the discipline of source selection and evaluation is the bridge between them.
Retrieval is the usual route to grounding, as in retrieval-augmented generation, and the work of Gao and colleagues on citations evaluates how well model answers cite the passages that support them. Grounding can also be to structured records, to tool results or to a user-provided document. In each case the test is the same: what is the supporting material, and does it say this?
Degrees of grounding and the risks of looking grounded
Grounding is not binary. A statement can be fully supported, partly supported, supported only by inference, unsupported or contradicted. The NIST profile on generative AI, AI 600-1, names confabulation, the production of confidently stated but erroneous content, as a risk to be managed, and it is the appearance of grounding that makes it dangerous: a fluent sentence with a citation looks checked even when it is not. Audits of generative search engines, such as the verifiability study discussed under sentence-level citation, found that many cited statements were not fully supported by their citations.
Measure grounding at the level of the claim. Break an answer into its claims, find the passage offered for each, and judge support. Automatic judges can scale this up, provided they are calibrated against human judgments, as described in the article on an LLM judge groundedness calibration protocol. Report the share of claims supported, partly supported and unsupported, and read the unsupported cases for patterns.
A worked example
An answer about Halden Therapeutics AG (fictional) says: the company completed its financing round in June, raising an amount that the answer states, and has since doubled its headcount. The first clause cites a press release that supports the completion and the amount. The second clause cites the same release, which says nothing about headcount. A document-level check passes the answer, because every claim has a citation to a real and relevant page. A claim-level check finds one supported claim and one unsupported claim. The second may be true, as an inference from a different source, but it is not grounded in the cited material, and the answer is either corrected, given a proper source, or the claim is removed. The audit tally records one of two claims grounded.
Checks a reviewer can run
Sample answers and decompose them into claims. For each claim, open the cited passage and decide whether it supports the claim fully, partly or not at all. Compute the grounded share. Look for statements with no citation at all. Check that the cited passages are retained, so the check can be repeated later. Compare the human tally with any automated judge's.
Related terms and boundaries
Grounding is realised by sentence-level citations to an evidence span, measured as faithfulness, and its absence is hallucination. It depends on the source reliability of what is cited, and it is achieved in practice through retrieval-augmented generation.
Common misreadings
- Using a citation that mentions the topic but does not support the claim.
- Hiding a conflict by citing only one source.
- Treating source support as proof of source reliability.
Related terms
Appears in
- Long context vs retrieval
Long context places more material in one model input; retrieval selects a smaller set of material for each question.
- Retrieval-augmented generation vs a system of record
Retrieval supplies context for an answer; a system of record governs which values may be treated as settled.
- Auditing an LLM answer against its sources
Split an answer into claims and inspect cited passages for support, inference and missing qualifications.
References
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS (2020).
- Rashkin et al. (2021), Measuring Attribution in Natural Language Generation Models, Computational Linguistics.
- Gao et al. (2023), Enabling Large Language Models to Generate Text with Citations, EMNLP.
- NIST (2024), AI 600-1: Artificial Intelligence Risk Management Framework, Generative AI Profile.
- Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS.
- Shuster et al. (2021), Retrieval Augmentation Reduces Hallucination in Conversation.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Grounding" bear on a decision of yours?
This entry sits under context engineering. If a decision in your organization turns on it, send the entity concerned, the sources you fuse to describe it and who may read which of them. We reply with what a first record would cover.
Or write to [email protected] in your own words.