AI / LLM · 2026
ScholarRAG: Question Answering over Scientific Papers
A retrieval-augmented system that answers questions about 200 arXiv papers on visual anomaly detection. Every answer has to quote the passage it comes from, the quote is checked in code, and the system declines when the papers do not contain the answer. Retrieval combines BM25 and dense search with a cross-encoder reranker, and Gemini writes the answer. On frozen test sets, the right passage is among the top 5 for 91–93% of questions, 92–93% of answers are judged acceptable with none wrong, and all 20 near-miss questions are refused.