Verified Research RAG

Ask a question over a 250-paper arXiv corpus (NLP / IR / LLM / ML-systems). The system retrieves evidence, generates a cited answer, then verifies every claim against the evidence it cites with a fine-tuned faithfulness verifier — and reports how much of its own answer is actually grounded.

🟢 Supported · 🟡 Weak · 🔴 Unsupported · ⚪ Abstention (the model correctly said the sources don't cover it) — try the third example: it asks about something the corpus doesn't cover.

Honest limitations. Generation is a hosted LLM call — the verifier is the contribution, not the generator. The verifier (DeBERTa, fine-tuned on a leakage-safe SciFact+HealthVer split) is recall-oriented and imperfectly calibrated: scores are useful as labels/rankings, not literal probabilities. A rigorous in-domain adaptation study produced a negative result — see the repo write-up.