Abstract

This study will entertain the design and implementation of a generative AI-based information retrieval system to enhance university documentation access. On the contrary, other search engines that look for keywords and provide irrelevant or lacking results, the system hence uses semantic search and natural language interaction with LangChain, Facebook AI Similarity Search (FAISS), and the OpenAI API integrated. Document embedding is used to place documents in high-dimensional vector spaces, while Retrieval-Augmented Generation (RAG) helps provide responses that are context-aware, accurate, and in natural language. The architecture comprises Django in the back-end and Angular for the front-end, designed to scale and remain modular to handle the processing of complex, poorly indexed, and unstructured academic documents with ease. University documents are processed through OCR-enabled pipelines, chunked, and stored in FAISS for fast similarity-based matching. LangChain then connects the retrieval results to the generative model, guaranteeing that the response is well-grounded on actual documents, thereby circumventing hallucinations and providing human-sounding fluency. While the evaluation provided ample evidence about the efficacy and accuracy of the approach, an overall accuracy of 89% was recorded, while, for specific questions, it was as high as 92%. The average time taken to retrieve was 1.33 seconds, and generation took around 3.3 seconds, all of which had user ratings above average for categories like relevance, ease of use, and satisfaction. It is, therefore, clear that vector-based semantic retrieval coupled with generative AI offers a promising alternative to circumvent the shortcomings of traditional search methods. Emerging problems, however, include embedding and inference costs, dependency on third-party APIs, and difficulties in text extraction from scanned or locally formatted, very untidy PDFs. Future work includes performance improvements such as caching and incremental indexing, better OCR, and better multilingual and adaptive learning support. With these characteristics, the system can pursue aggressively with the implementation of its functions in the academic environment, thereby bridging the gap at the two ends of keyword search and actual context-based information access.