Asistente inteligente para consulta asistida por IA y recuperación de información en documentos PDF académicos

Abstract

The objective of this work was to develop and locally verify an intelligent assistant to query academic documents from a repository of the Catholic University of Cuenca. In this context, the problem is that lengthy documents make it difficult to quickly and accurately locate information when relying solely on traditional keyword search. To address this, we implemented a prototype that converts content into vector representations, performs semantic search using the Facebook AI Similarity Search (FAISS) library, and generates locally executed responses with Ollama. The prototype has a client-server architecture and integrates document loading, text chunking, semantic retrieval, and history persistence through Retrieval Augmented Generation (RAG). Additionally, the prototype incorporates references to the analyzed document and clearly indicates when there is insufficient evidence to answer. The methodology used was Scrum, and we conducted end-to-end tests on single-file and batch loading. We also applied a perception survey to four participants to assess ease of use, clarity, usefulness, and user experience. Regarding performance results, the current version of the prototype increases document ingestion and indexing time compared to the initial version. This is because the second version generates more embeddings but improves evidence retrieval and response quality. It is concluded that an intelligent document assistant optimizes the query process by reducing search times and eliminating dependence on external services.