Desarrollo de un asistente virtual inteligente basado en arquitectura RAG para la gestión de información académica y logística en la Facultad de Ingeniería en la Carrera de Ciencias de la Computación.

Abstract

Academic and logistical information for the Facultad de Ingeniería en la Carrera de Ciencias de la Computación at the Universidad Católica de Santiago de Guayaquil (UCSG) is spread across several official channels: the academic portal, curriculum documents, the institution's Instagram account, and the front desk staff's knowledge of the physical location of classrooms and laboratories. Each channel serves its purpose, but a student looking for a specific answer needs to know beforehand which channel holds it. This work presents the development of Compu IA, a conversational virtual assistant accessible through WhatsApp, built to centralize that lookup in a single interface.

The system combines a hybrid architecture composed of PostgreSQL with the pgvector extension, a FastAPI orchestrator, a web administrative panel built with Next.js and TypeScript, a local inference engine based on Ollama under an optional profile, and automated information extraction processes, with a five-level inference cascade router (Groq, Gemini, Claude, GPU-accelerated Ollama, and CPU-based Ollama), which keeps the service operational when one of the artificial intelligence providers fails or runs out of quota. Information retrieval was resolved through a Text-to-SQL architecture rather than through semantic vector similarity search, in order to prioritize the accuracy of the data delivered over the flexibility of the search. On top of that retrieval layer, a set of deterministic validations was implemented (regular expression filters, structural verification of the generated SQL, and a mechanism that discards any model response contradicting the actual retrieved data), reducing the system's dependence on the language model's behavior for correctness. The economic viability of the architecture was also evaluated, comparing the estimated monthly cost of cloud operation against the acquisition and maintenance cost of dedicated on-premise GPU infrastructure, and the prototype's limitations and future work were documented.