https://www.mdu.se/

Manufacturing industries rely on vast collections of multi-modal documents for

product development and maintenance, encompassing hardware specifications, soft-

ware documentation, and technical diagrams across diverse formats. Efficiently

retrieving relevant information from these complex documents presents significant

challenges due to their domain-specific terminology and structural complexity. This

thesis investigates the application of Retrieval-Augmented Generation (RAG) sys-

tems with Large Language Models (LLMs) for manufacturing document search and

question answering. The research compares traditional vector-based RAG approaches

with advanced graph-based methods like LightRAG to evaluate their effectiveness for

industrial documentation retrieval. A comprehensive preprocessing pipeline was de-

veloped to handle multi-modal content, extracting structured information from text,

tables, and technical diagrams while preserving document context. Experimental

evaluations using documents from the European Union Agency for Railways demon-

strate that different RAG architectures excel in different scenarios: vector-based

approaches with advanced prompting strategies performed well for specific low-level

queries, while graph-based global retrieval strategies showed superior performance for

complex questions requiring synthesis across multiple documents. While automated

metrics showed advanced prompting strategies achieving higher ROUGE and BLEU

scores, manual analysis revealed that graph-based methods often produced more

comprehensive and contextually relevant answers for complex queries. This research

contributes to the understanding of RAG systems for industrial applications and

provides insights for optimizing multi-modal document retrieval in manufacturing

contexts.