swarajsonawane4/scifact-rag

SciFact RAG + Groundedness Probe

End-to-end Retrieval-Augmented Generation pipeline on BEIR SciFact with sentence-level evidence probing.

Pipeline

1. 1.Dense retrieval: Token-chunked corpus → MiniLM embeddings → FAISS cosine search
2. 2.Cross-encoder reranking: MS-MARCO MiniLM rescores top candidates
3. 3.Sentence-level NLI probe: DistilBART-MNLI scores each sentence for entailment/contradiction
4. 4.Keyword-overlap gating: Filters off-topic NLI false positives
5. 5.Conservative verdict: SUPPORTED / REFUTED / INSUFFICIENT with citations

API Endpoints

Metrics (from evaluation notebook)

• —Dense-only: Recall@10 ≈ 0.80, MRR@10 ≈ 0.61
• —Dense + rerank: Recall@10 ≈ 0.83, MRR@10 ≈ 0.67

Built by

Sudhanshu Pawar — M.S. Applied Data Science, Syracuse University