Groundedness & attribution checking

How it works

Groundedness asks a narrower question than factual correctness: not 'is this true?' but 'is this entailed by the context the system itself retrieved?' The usual shape is to split the response into atomic claims, then test each against the supplied passages — by entailment model, by judge model, or by requiring an explicit citation that resolves. An ungrounded claim is a hallucination that retrieval was supposed to prevent, and it is the failure mode a RAG system exists to eliminate. The same machinery scores attribution: whether cited passages genuinely support the sentence they're attached to, rather than merely being on-topic.

When to use it

Any system that answers from retrieved or supplied documents — RAG, document Q&A, summarization from source material; auditing citation quality before publishing generated content; setting a release bar for RAG faithfulness that a benchmark accuracy score can't express, since a groundedness failure can hide behind an answer that happens to be true anyway.

Limitations

A perfectly grounded answer can still be wrong when the retrieved source is wrong — groundedness measures faithfulness to context, not truth, so it's not a substitute for factual-oracle verification when a trustworthy oracle exists. Claim decomposition and entailment judging are themselves model-driven and inherit those models' errors, so calibrate against human labels before trusting a threshold.

Cite this