Earthquaker-AI: A Retrieval-Augmented Generation Framework with Rubric-Based Assessment for Primary School Earthquake Education

TL;DR — Earthquaker-AI is a hybrid educational framework that adds a Retrieval-Augmented Generation (RAG) conversational assistant to an existing Lego WeDo2 earthquake-preparedness robotics project, aiming to teach primary-school students safe crisis-management actions. It matters as an applied example of grounding an LLM assistant in official safety guidelines to deliver reliable, age-appropriate tutoring.

• RAG-based dialogic module: Student queries are semantically matched against official earthquake-safety guidelines to generate grounded, accurate responses; reported evaluation shows high groundedness/accuracy and a low hallucination rate (specific metrics not provided).
• Rubric-based, developmentally staged assessment: Feedback scales by grade — 2-D rubric with multiple-choice recognition (early), 3-D rubric for action-sequence identification (middle), and 4-D rubric evaluating short written responses including clarity of expression (upper).
• Hybrid hands-on + cognitive design: Combines tangible robotics (sensors/actuators simulating seismic response) with AI-guided verbal feedback to support self-regulated learning and calmness under simulated emergencies.
• Note: This is an applied educational-systems paper; it emphasizes pedagogy and responsible AI use rather than novel model architecture or benchmarked quantitative results.