Development of a Semantic Retrieval System Using Retrieval-Augmented Generation (RAG), GPT Models, and Vector Databases to Enhance Academic Knowledge Search and Analysis

Abstract

     This study developed a semantic retrieval system for academic knowledge by integrating the Retrieval-Augmented Generation (RAG) architecture with GPT models and vector databases to support natural language queries. Adopted a software engineering framework, the research involved analyzing existing system limitations, designing, implementing, and evaluating the RAG-based system. The system successfully ingested 96.21% of research documents into vector representations. Evaluation of summarization quality showed an average Compression Ratio of 0.74, Cosine Similarity of 0.9597, and ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.65, 0.56, and 0.53, respectively. For retrieval performance, expert evaluation revealed that Semantic Search achieved 90% Precision and Recall and a 98% Mean Reciprocal Rank (MRR), compared with 79% Precision/Recall and 86% MRR for the Keyword Search baseline. Overall, integrating RAG, GPT, and vector databases enhances accuracy, relevance, and usability in academic information retrieval through natural language interaction, demonstrating strong potential for expansion to large-scale repositories.