Design of a Portable Retrieval-Augmented Generation Architecture for Offline Intelligent Tutoring on Resource-Constrained Hardware
Year: 2026
Type: article
Abstract: Cloud infrastructure dependency limits the adoption of intelligent tutoring systems in connectivity-restricted contexts, affecting approximately 2.6 billion people without internet access. Objective: This article presents the design of a portable intelligent tutor that operates fully offline from a USB device on mid-range hardware without a dedicated graphics processing unit. Methodology: The design was approached through Design Science Research (DSR), structured in three iterative cycles covering component selection, architectural integration, and quality evaluator tuning. Results: A reference architecture is proposed that integrates hybrid retrieval (dense vector search, BM25 lexical retrieval, and Reciprocal Rank Fusion), local inference with quantized language models in GGUF format, a pedagogical prompt builder with four teaching modes, an automatic quality evaluator, and AES-256-GCM encryption at rest. The validation, technical and preliminary in scope, consisted of running the system from the same USB drive on three machines with heterogeneous hardware profiles: it operated functionally on a computer with a dedicated GPU (latencies of 54 to 111 seconds per query) and on a mid-range laptop without a GPU (7.7 GB of RAM, 2,048-token context window), while a third machine lacking AVX instructions exhibited an execution failure that delimits the compatible hardware floor. Conclusions: The results suggest a technically viable pathway for reducing access barriers to AI-assisted tutoring in contexts with limited technological resources and connectivity; its pedagogical effectiveness remains to be evaluated with end users.
Source: ENTRAMADO
Institution: Institución Universitaria Colegio Mayor del Cauca
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7211878552