Abstract

Recent advancements in Large Language Models (LLMs) have transformed Natural Language Processing, yet challenges such as factual inaccuracies and inadequate reasoning over structured data persist. Retrieval-Augmented Generation (RAG) systems address these issues by grounding LLMs in external knowledge. However, conventional RAG methods typically treat knowledge sources as unstructured text, overlooking the semantic relationships vital in domains like enterprise data and healthcare. This thesis introduces a Graph-based RAG approach that leverages the structured nature of graph data to enhance both retrieval and response generation by preserving these semantic relationships. The research focuses on developing conversational question answering systems over semantically structured data, specifically targeting JIRA Issues and Knowledge Graphs through two distinct applications. The core innovation lies in maintaining the inherent data relationships during both retrieval and generation phases by employing structured queries and graph traversal techniques. This method not only allows for domain-specific optimization but also demonstrates improved performance and efficiency compared to traditional RAG methods.