PracticeEvaluation & Safety
Groundedness (Faithfulness)
At a glance
Whether every factual claim in an answer is supported by the passages that were retrieved.
- Who this is for
- Engineers and technical readers learning the terms used in AI systems.
- Topics
- Evaluation & Safety
- Practice
In short
Groundedness asks whether every factual claim in an answer is supported by the retrieved passages, the same check Ragas calls faithfulness. Bigcircle scores claims with a binary or three-way rubric and tracks citation validity and completeness as separate numbers. Our guide on document intelligence beyond extraction keeps a page and span on every value, and treats a figure with no path back to a page as unsupported.
Key facts
Definition
Groundedness is the check that each factual claim in an answer is supported by the passages retrieved for that answer, independent of whether the answer is complete or helpful.
How Bigcircle does it
- Bigcircle scores each claim against the passages with a binary or three-way rubric, because careful readers usually agree on support and disagree on whether an answer is good.
- Track citation validity on its own, because a real clause can still be the wrong clause, and track completeness on its own, because one supported fact can omit the rest.
- Store lineage to the span when a regulated reader must trust the output, because a footnote to the whole document does not let review check the cited place.
Comparison
Groundedness asks only whether some retrieved passage supports the claim, citation validity asks whether the cited clause governs the question, and completeness asks whether other needed facts were left out.
Groundedness asks one question: is every factual claim in this answer supported by the passages that were retrieved? Ragas calls the same measure faithfulness, and the two names refer to the same check.
It matters because retrieval and generation fail separately. The right passage can reach the model and the model can still assert something the passage does not say, which is a different defect from a retrieval miss and needs a different fix.
Why it is checkable when "good" is not#
Most generation-side quality questions are too vague to score consistently. Two reviewers will disagree about whether an answer is helpful, and a model judge asked the same thing will reliably prefer the longer one.
Groundedness avoids that because it is close to mechanical. Take each claim, look for it in the passages, decide whether it is there. Two careful people will usually agree, which is the property that makes a check worth automating.
What it does not cover#
Groundedness says nothing about whether the answer is complete. A response that states one supported fact and omits three others is perfectly grounded and still wrong for the reader, which is why completeness is tracked separately.
It also says nothing about whether the passage was the right one. An answer can faithfully quote a clause that does not govern the question asked. That is citation validity, and on legal work it is the more dangerous failure: a correct number pulled from the wrong clause survives a reviewer's scan in a way an obviously invented one does not.
Citations as architecture#
Where a regulated reader has to trust the output, groundedness has to be provable rather than asserted, which means lineage to the span and not a footnote to a document.
On our ESG filings platform every extracted metric carries lineage to the page. That moved the analysis cycle from weeks to hours, because review stopped being "re-read the filing" and became "check the cited page".
Practical takeaway#
Score groundedness with a binary or three-way rubric rather than a scale, ask about claims rather than about quality, and track citation validity and completeness as separate numbers. Together they answer whether the answer is supported, whether it is supported by the right thing, and whether anything was left out.