Published Online:June 2025
Product Name:The IUP Journal of Information Technology
Product Type:Article
Product Code:IJIT020625
DOI:10.71329/IUPJIT/2025.21.2.24-43
Author Name:W P K I Perera, P Vigneshwaran and J Charles
Availability:YES
Subject/Domain:Engineering
Download Format:PDF
Pages:24-43
The paper investigates the evolution of retrieval augmented generation (RAG) architectures, critically examining current solutions and their inherent drawbacks in handling information retrieval, document chunking, utilization of updated information in domain-specific tasks, questionand-answering, and reasoning applications. The review addresses four primary research questions: mechanisms proposed in literature for mitigating hallucinations and enhancing private data accessibility; evaluation of existing RAG solutions and their identified limitations; strategies documented for improving RAG architecture to extract more relevant information; and comprehensive performance assessment methodologies employed in current research. The review analyzed 20 carefully selected papers published between 2020 and 2024, applying stringent inclusion and exclusion criteria. The key findings reveal significant innovations documented in RAG approaches, including schema-less knowledge graph development, advanced retrieval mechanisms, and multi-agent feedback loops. The review synthesizes research demonstrating the underutilization of techniques such as personalized PageRank algorithms for retrieval, iterative query reformulation, and external knowledge verification tools that studies indicate demonstrate potential for enhancing RAG system accuracy and reliability
Retrieval augmented generation (RAG) has emerged as a promising advancement in addressing several inherent challenges associated with large language models (LLMs). These challenges include hallucinations, where models generate inaccurate or nonsensical information, and limitations in accessing private or sensitive data and difficulties in processing complex sets of instructions.