Enhancing Retrieval Augmented Generation Systems with Knowledge Graphs

Knowledge graphs play a pivotal role in information retrieval, yet their effectiveness can be further optimized. This paper introduces a comprehensive approach to enriching knowledge graphs, addressing contemporary challenges in answering user queries. Our methodology integrates key phrase extraction indulged with a custom prompt to create a Knowledge Graph (KG) with every node connected with each other, node embedding generation w.r.t properties, and an autonomous updating agent during the creation of KG and inference. Furthermore, we explore the incorporation of traditional vector search to enhance contextual understanding. Our experiments demonstrate a substantial improvement in accuracy, reaching approximately 96% compared to traditional KG approaches and our creation of KG process. Notably, the hybrid model sometimes outperforms the Retrieval-Augmented Generation (RAG) system, showcasing the efficacy of the integrated approach. This enhancement is attributed to the methodology proposed coupled with the additional context provided by traditional vector search. The results underscore the significance of our approach in delivering more accurate and contextually relevant information, showcasing the potential of this integrated method in advancing knowledge graph systems.