Loading page
About this tool
Check locally whether answer claims carry recognizable citations to the supplied retrieval set, without claiming that a source proves the claim.
The RAG Citation Coverage Checker splits a generated answer into claim sentences and reports, for each one, whether it carries a citation marker that resolves to an ID in the retrieval set you paste alongside it. Every claim comes back as linked, mixed, unknown or missing, and the coverage figure is the share of claims carrying at least one recognised source ID, to one decimal place. This is a structural check for people evaluating RAG pipelines: it tells you whether a citation points at a real retrieved chunk, never whether that chunk actually supports the sentence.
Open RAG Citation Coverage Checker on AltFTool — it loads instantly in your browser.
Paste the answer into Generated answer and the retrieval set into Retrieved sources — JSON, or one line per source as ID | passage — or use Open text for a .json or .txt file up to 1 MB.
Press Check coverage to fill the Coverage, Claims, Linked, No citation, Unknown only and Sources tiles.
Work down the Claim map, where each sentence reads Linked, Mixed references, Unknown references or No citation, then use Counts-only report to download rag-citation-coverage-report.json.
A claim is linked when every citation resolves, mixed when some resolve and some do not, unknown when none resolve, and missing when there is no citation at all.
Paste JSON - an array, an object keyed by ID, or a {sources: [...]} wrapper - or plain lines like "S3 | passage text", and IDs are matched case-insensitively with surrounding brackets stripped.
Headings, questions, a trailing "Sources:" block and fragments under four words are excluded outright, and leading list-marker syntax (bullets, numbering) is stripped from the remaining text, so the coverage denominator stays honest.
It is the number of claims with at least one recognised citation - the linked and mixed ones - divided by the total number of claim sentences, rounded to one decimal place. Claims whose only citations are unrecognised IDs count against you.
Square-bracket markers such as [3], [doc-12] or [a, b] with comma, semicolon or "and" separators, and parenthetical labels such as (source: S2) or (refs: 4, 5). An ID may be up to 160 characters of letters, digits and the separators _ . : / -
No. The check is purely structural - it confirms that a citation marker exists and matches an ID in the retrieval set you supplied. It cannot tell you the passage is relevant, the claim is true, or that the retrieval set was complete; that still needs human or model-based grading.
Up to 1,000,000 characters per input, 5,000 sources and 5,000 claims, with each source passage read to 100,000 characters. If anything is cut off, the result is flagged as truncated so you know the numbers are partial.