Eval bench · 001

RAGJudge

Prove your RAG works. Feed it documents and test questions, pit retrieval configurations against each other, and score every contestant on faithfulness, citation quality, and answer relevance — all in your browser, no API key required.

embeddings: checking… answer mode: extractive (no key)

1Dataset

Documents to retrieve from, plus test questions with reference answers and gold sections. A fictional sample dataset is bundled — hit Load sample and skip to step 2.

Documents 0

+ Add a document (paste text)

Test questions 0

+ Add a test question

2Config matrix

Each combination becomes one contestant. Cosine retrieval needs the embedding model (downloads ~90MB once); without it, cosine configs fall back to BM25 and say so.

Contestants: 0

4LLM judge (optional, BYOK)

Re-scores faithfulness 1–5 with a rationale, shown side-by-side with the heuristic score — it never replaces it. Runs against the latest leaderboard.

RAGJudge runs 100% in your browser — documents, keys, and results never leave this device. Heuristic scores are token-overlap based, not NLI: see README for metric definitions and honest limitations. All scorers are original implementations.