Catch the numbers your RAG answer made up
Paste an answer and the sources it was given. groundcheck flags figures that appear in no source, claims nothing supports, citations that point at the wrong source and negations that got flipped. It is deterministic: the same input always gets the same verdict, so it can gate a CI pipeline where a model acting as judge cannot.
Everything runs in this page. Nothing you paste is sent anywhere, and no model is called.
Loading Python in your browser. About 10 MB the first time, cached after that.
Sentence by sentence
Each claim, whether a source supports it, and how much of its wording the closest source shares. Connective sentences with nothing to check are left out.
The same check in your pipeline
pip install git+https://github.com/RootX22/groundcheck.git
groundcheck --answer answer.txt --source doc1.txt --source doc2.txt
echo '{"answer": "...", "sources": ["..."]}' | groundcheck --json-input - --fail-on highStandard library only, exit codes for CI, MIT licensed. Source, docs and the limits of the heuristic.