[2025S] Andrew Chung (BS)

Low-Resource RAG

Andrew Chung

Date: 2025-02-14 / 2:00 - 3:00 PM
Location: White Hall 100

Abstract

Large Language Models (LLMs) have become essential across education, research, and business, particularly for providing domain-specific information. Retrieval-Augmented Generation (RAG) has emerged as the leading framework for leveraging LLMs while maintaining accuracy on domain-specific information. We examine RAG components to identify essential elements and improvements. Through collaboration with Hyundai, we develop a low-resource domain RAG system for automotive safety collision tests using multimodal slides. Our novel, language model-centric data processing pipeline effectively transforms slide information into textual content suitable for retrieval and answer generation. We attempt to open-sourc this pipeline by fine-tuning Qwen 2.5 VL 7B, with the potential to one-shot the entire data processing pipeline. We evaluate different RAG frameworks and embedding models using synthetic question-answer pairs, including pairs targeting multimodal table and chart information. Results show that fine-tuned embedding models with the original RAG framework achieve highest accuracy. We outline next steps for developing open-source RAG frameworks for low-resource domains.

Link

Presentation