Polp
← Back to glossary
AI quality and safety

Grounding

The relationship between an AI output and the supplied evidence, where each material claim should be supported by the available context or facts.

Business example

An assistant answers an expense question from the current travel policy and links each limit to the supporting clause. A claim not present in that policy is treated as ungrounded.

How it works

  1. 1.Retrieve or supply the reference passages allowed for the user and question.
  2. 2.Separate the answer into factual claims that require support.
  3. 3.Check whether each claim follows from one or more reference passages.
  4. 4.Return citations, support signals, or an abstention when the evidence is insufficient.
  5. 5.Evaluate relevance and completeness separately because a grounded answer can still miss the question.

Common misconceptions

A citation automatically proves that an answer is grounded.
The cited passage must actually support the specific claim; a related document or title is not enough.
Groundedness and relevance are the same metric.
An answer can repeat a true source fact that does not answer the question, or answer the topic with a claim absent from the source.

Related concepts

Retrieval-Augmented Generation (RAG)AI hallucinationKnowledge baseReranking

Related reading

The new AI SaaS metric: answer quality, not document count
How to evaluate internal AI search tools: a 15-question checklist

Sources

Reviewed:

Reviewed by: Javier Chulvi Bernad · LLM Engineer · Madrid