TokenPad
Evaluation

RAG Grounding Checker

Find the sentence the model added that the context never contained.

Context, then === , then the answer

255 characters5 lines0 tokensor drop a file

Grounding CheckerExact
0Token change
Token change0no change
Input tokens0what you pasted
Output tokens0what you would send
Token cost of this result
Output tokens0
As input$0.00
× 100K requests$0.00

Everything on this page runs in your browser. Nothing you paste is transmitted, because there is no server here to transmit it to.

Result

Fabricated numbers are the highest-signal defect

A figure in the answer that never appeared in the context is almost always invented. Models produce plausible numbers readily, and a plausible number is far more damaging than a vague sentence because people act on it.

The checker extracts every number from the answer and verifies it appears in the source. This is a cheap check with a high hit rate.

What this is not

It measures lexical overlap. A correctly paraphrased claim scores low and a fluent fabrication reusing context vocabulary scores high. It cannot determine truth.

Treat the output as a list of sentences worth a human reading, not as a verdict. Used that way it reliably surfaces the claim that slipped in from the model rather than the source.

Frequently asked questions

How do I reduce ungrounded claims at the source?
Instruct explicitly that claims not supported by the provided context must be declined, and require a citation per claim. Both measurably reduce it, though neither eliminates it.
Should I run this in production?
The numeric check, yes — it is fast, deterministic and catches the worst failures. Route anything flagged to review rather than blocking it, since the false positive rate on paraphrase is high.

More evaluation tools