Faithfulness and Relevance in Stable Normative Domains: A Comparative Analysis of CAG vs. RAG

Collanqui E.A.; Coaguila M.E.; Flores A.; Apaza H. · Communications in Computer and Information Science (2026) · DOI: 10.1007/978-3-032-20322-9_10

Resumen

In higher education, efficient applicant support is crucial but often limited by manual processes and the inherent constraints of large language models (LLMs) regarding factual accuracy within specialized domains. This study conducts a rigorous comparative analysis of Retrieval-Augmented Generation (RAG) and Cache-Augmented Generation (CAG) architectures to determine the optimal approach for an AI-driven chatbot in the university admissions domain. Using the university’s 2025 admission regulation as the knowledge base, we evaluated state-of-the-art LLMs, including Google’s Gemini (2.5 Flash & 2.5 Pro) and Meta’s Llama 3.1 (8B & 70B). Performance was assessed on an expanded dataset of 134 questions using the RAGAs framework, focusing on key metrics like faithfulness, answer relevancy, and response latency. Our results reveal that the CAG architecture generally achieves superior factual accuracy. Notably, the CAG + Gemini 2.5 Flash combination emerged as the optimal configuration in terms of quality, delivering the highest faithfulness (0.970) and answer relevancy (0.757). Crucially, the RAG configuration of this same model registered the lowest response latency, highlighting a clear trade-off between maximal quality and speed. This research provides an evidence-based blueprint for developing highly accurate and efficient AI chatbots for normative domains, demonstrating that a well-designed architecture with a cost-effective model can deliver state-of-the-art performance. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Instituciones: Universidad Nacional de Moquegua