MMOOZZGG.com

Glossary

What is groundedness in AI?

In short

Groundedness is a property of an AI answer: each of its claims is supported by the documents that were passed to the model for the answer.

Search and RAG · groundedness

MMOOZZGG.com editorsChecked: 1 min read

For a child

You tell the class about volcanoes and say: “That is what the encyclopedia says on page ten.” The teacher opens the book and sees it is true. So your story is grounded. If the book does not say it, the story is made up. Groundedness shows whether the AI took its words from the book or invented them.

In plain words

When an AI assistant answers from company documents, it is not enough for the answer to sound plausible. Each claim in it must be found in the documents the system passed to it. This is called groundedness. The check works like this: you split the answer into separate claims and look for support for each one in the retrieved text. If nine out of ten are supported, groundedness is ninety percent. For example, the assistant wrote that a return is possible within fourteen days, but the policy says seven: the answer is not grounded, even if it sounds confident.

For business

For a company, it is the main measure of trust in an assistant: an ungrounded answer to a customer about terms, prices or contract conditions creates legal and reputation risk. A second checking model can compute the metric automatically, without manual labels for each answer, so you can track it all the time, not only at launch. Know the limit: a grounded answer is not always correct. If the knowledge base holds an outdated policy, the assistant will repeat the error faithfully. So measure groundedness together with the quality of the base and the share of questions where the right document was found. Agree in advance on the threshold below which an answer goes to a person, and test these rules in a pilot.

For a specialist

Groundedness, also called faithfulness in the RAG evaluation literature, means that the claims of an answer can be inferred from the given context. It is not the same as factual accuracy about the world. In the RAGAS framework (Es et al., 2023), a language model splits the answer into short statements S and checks for each whether it follows from the context. The score equals |V|/|S|, where V is the set of supported statements. The same framework defines answer relevance and context relevance, and all three metrics are reference-free: they need no human-labeled reference answers. Groundedness is linked to intrinsic hallucinations, which contradict the source context. Ways to raise it: limit the answer to the context in the instruction, require citations, use Self-RAG-style self-checks where the model marks whether a retrieved passage supports a segment, and filter retrieval results as in Corrective RAG. An LLM judge makes mistakes too, so it is calibrated on a manually labeled sample.

Example

An assistant tells a customer that delivery is free for orders over 5,000. The check finds a different sum in the passed fragment, marks the answer as ungrounded and hands the dialog to an operator.

Related terms

Questions and answers

How is groundedness different from a correct answer? Groundedness means the answer relies on the documents passed to the model. Correctness means it is true in substance.

If a document is out of date, a grounded answer will still be wrong.

How do you measure the groundedness of answers? Split the answer into claims and check for each whether the retrieved text supports it. A second model does this, and people spot-check its scores.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us