Grounding — AI Glossary

Reference details

Related terms

• RAG (Retrieval-Augmented Generation)
• Hallucination
• Function Calling

How it works

1. Evidence is obtained before generation rather than recalled during it — retrieved passages, a tool result, a database row.
2. That evidence enters the prompt in a position the model is instructed to treat as the source of truth for the answer.
3. The instruction constrains the output to what the evidence supports, including an explicit path for abstaining when it does not support anything.
4. Each claim is emitted with a pointer to the span it came from, which is the part that makes the answer checkable rather than merely sourced.
5. A verification pass can then compare claims against the cited spans — the step that catches an answer which cites real evidence and still asserts something the evidence does not say.

Example: Citing evidence is not the same as being supported by it

A system retrieves a document stating a service had 99.9% availability in Q3 and answers: 'Availability was 99.9%, the best of any quarter.' The citation is real and the first clause is supported. The second is not — the document covers one quarter and cannot establish a comparison across quarters. Grounding failed at the claim level while looking correct at the document level, which is why per-claim verification is a separate step from retrieval and why an answer with a footnote is not thereby grounded.

Commonly confused with

Treated as a synonym for RAG, which is one technique for achieving it. Grounding is the property — output constrained to verifiable evidence — and retrieval, function calling and schema-constrained generation are all ways to get there; a tool returning a live figure grounds an answer with no retrieval involved. The second confusion is with hallucination reduction: grounding reduces unsupported claims, but a model can ground a false statement in a false source, so grounding makes an answer checkable rather than true. The check is what adds the truth, and it is a step people skip.

When to use it

Reach for it when:

• Answers concern facts that change, or that live in a corpus the model never saw.
• The user must be able to verify a claim, not just receive it.
• Abstention is preferable to a plausible guess, which requires evidence to abstain against.

Reach for something else when:

• Tasks with no external truth to ground against, such as drafting, rewriting or ideation.
• As a hallucination guarantee: it constrains and makes claims checkable, and does not verify them.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Grounding — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path