INTERNATIONAL JOURNAL FOR INNOVATIVE RESEARCH IN MULTIDISCIPLINARY FIELD

Abstract: Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems are increasingly used to answer questions from scientific literature. However, generated answers may contain unsupported claims or conflicting information, making verification difficult. This paper presents EvidenceGraph, a framework that links individual claims to relevant evidence and source documents. It combines semantic retrieval using ChromaDB, answer generation, claim extraction, evidence verification, and graph-based relationship analysis. Evaluated on 30 questions from three scientific papers, the system produced 96 claims, of which 62.5% were supported, 20.8% contradicted, and 16.7% had insufficient evidence. Target-paper recall reached 100% at Top-10 retrieval, and all 151 automated tests passed. The results highlight EvidenceGraph’s potential for improving evidence traceability and claim-level auditing in scientific RAG systems.