Laya RAG Judge

Retrieve evidence, then let a typed-decision model judge the claim, in your browser

Models not loaded

Judge: . Retriever: . Both download once and are cached by your browser. Until then, the built-in claims play back results recorded from the same models.

Pick a sample claim or write your own, then press Verify claim.

How it works

Every document is split into short passages and embedded with all-MiniLM-L6-v2; the vectors live in this browser's IndexedDB. The claim is embedded the same way and a brute-force cosine scan returns the top-k passages. The claim and those passages become the state of one typed question for the Laya decision model, which returns a probability for each option instead of generated text. That probability drives the verdict, and its confidence decides whether the verdict is acted on or held for a person. All of it runs in this tab: no server, no API key.

The Evaluate page runs the same pipeline over all 72 sample claims and scores it against the gold labels.