Concept dictionary
Grounding
Grounding is the practice of connecting an AI output to specified evidence, data, tool results, or current conditions that can be inspected.
- Review state
- Reviewed in POC
- Freshness
- Current
- Version
- 1
Also called
- Grounding
- AI Grounding
Remember it this wayGive the answer a basis it can point to.
In plain language
Instead of asking a model to answer only from patterns learned during training, a grounded system gives it a stated basis for the current response. That basis might be approved documents, database results, web sources, tool output, or the current state of an environment. A useful implementation also makes the basis visible enough for someone to inspect.
Why it matters
Grounding can make answers more current, relevant, and traceable while reducing unsupported invention. It matters most when the model needs private facts, changing information, or evidence that a learner or reviewer must be able to examine.
Example
An employee asks a benefits assistant whether a treatment is covered. The assistant searches the current policy library, answers from the matching plan document, and links to the exact section and effective date.
Common misconception
A grounded answer is not automatically correct. The source can be false, outdated, irrelevant, or misread by the model. Retrieval-Augmented Generation is one way to provide grounding; grounding supplies an inspectable basis, while verification checks whether the resulting claims hold up.
Related terms
Sources
- Google Cloud, “Grounding overview”defines grounding as connecting model output to verifiable information sources and describes traceability benefits.
- Google Cloud, “Check grounding with RAG”distinguishes a generated answer from the reference facts used to assess its support.
Learn more
- Compare the supporting material supplied through [[grounding|Grounding]] with the claim checking performed through [[verification|Verification]].
- Use [[evidence-selection|Evidence Selection]] to ask whether the sources chosen for grounding are fit for the question.