HAnS: Multimodal RAG-based Persona Generation for Media and Documents in E-Learning

Authors: Thomas Ranzenberger, Steffen Freisinger, Tobias Bocklet, Korbinian Riedhammer

Abstract:

We present an integration of personas into large language models (LLMs) within the HAnS learning experience platform. In HAnS, personas are used to tailor LLM responses to specific educational contexts and tasks. We use multimodal retrieval-augmented generation to provide contextually relevant responses for the learning content provided, including videos, podcasts, and supplementary materials. The platform enables lecturers to semi-automatically create personas based on learning materials using a template-based approach. They can also configure chat modes, context restrictions, and guard prompts to ensure appropriate behavior. The system enables the iterative testing and refinement of personas to address challenges such as prompt clarity and role consistency. In practice, this approach ensures personas behave consistently in live deployments, supporting diverse educational use cases.