16–21 Aug 2026
Daejeon Convention Center
Asia/Seoul timezone

Performance Comparison of LaTeX Preprocessing Methods and Embedding Models for RAG on RAON Control System Technical Notes

MOPO082
17 Aug 2026, 16:00
2h
1F Exhibition Hall (Daejeon Convention Center)

1F Exhibition Hall

Daejeon Convention Center

Poster Presentation MC4.A05: Other technology MOPO - Poster Session

Speaker

Won Hee Min (Institute for Basic Science, Chungnam National University)

Description

RAON uses EPICS as the primary framework for inte-grating local control systems into the central control sys-tem. Technical notes for control system installation and management are documented in LaTeX format and man-aged with relevant code and files in a Git-based configura-tion management environment on an isolated internal network. As part of efforts to introduce AI technologies into the control infrastructure, this study examines the applicability of Retrieval-Augmented Generation (RAG) to control system technical documents.
LaTeX documents were converted into formats such as Markdown under multiple preprocessing scenarios with different document representations and chunk segmenta-tion conditions. Different embedding models were applied to construct vector databases. Relevant queries were de-signed, and performance was compared in terms of re-trieval accuracy and contextual relevance. This study aims to evaluate the feasibility of applying RAG to EPICS technical documents and to suggest directions for build-ing a retrieval framework for LaTeX-based documentation. BGE-M3 with Structured LaTex and heading-based chunking achieved the best retrieval performance, with an MRR of 0.647 and Hit@5 of 0.777

I have read and accept the Privacy Policy Statement Yes

Author

Won Hee Min (Institute for Basic Science, Chungnam National University)

Co-authors

Prof. Jong-Ryul Lee (Chungnam National University) Dr hyun man jang (Institute for Basic Science)

Presentation materials

There are no materials yet.