Specialized RAG development for knowledge transfer using technical papers: A study on figure and table extraction methods

Abstract

Technical reports in the port and coastal engineering domain accumulate practical knowledge on the design and construction of structures and are therefore useful data sources for AI-assisted knowledge reuse. However, these reports rely heavily on non-textual evidence such as cross-sectional drawings, layout drawings, photographs, and specification tables, so text-centered retrieval alone often fails to surface relevant precedents. This study does not propose a new learning algorithm; rather, it designs and evaluates a domain-specific retrieval-augmented generation (RAG) workflow for standardized technical reports, focusing on knowledge representation, retrieval design, and task-based evaluation. The target documents were limited to materials whose use and transmission to the OpenAI API were permitted within the joint research, while direct reproduction of source text, figures, and tables in the paper was excluded. We compared four methods for extracting figures and tables from technical report PDFs and adopted a document-layout analysis model as an operationally suitable extractor. We then implemented three RAG variants that differ in how visual information is supplied: text-only retrieval, retrieval over text plus pre-generated figure/table captions, and question-time selection of nearby figures followed by captioning and answer synthesis. Search-space filtering based on structured metadata was also examined. The system was evaluated using practitioner-reviewed questions from three perspectives: accuracy, evidence consistency, and practical usefulness for design checks and responses to technical issues. Four repeated runs of the same LLM judge were used only to assess internal consistency, not as a substitute for human evaluation. The results indicate that incorporating pre-generated figure/table captions into retrieval improves answer quality for tasks requiring design-condition confirmation and interpretation of section-specific specifications. The proposed workflow shows the potential to support information retrieval and verification work for technical knowledge transfer, while further evaluation of retrieval-level metrics, caption quality, and fully on-premise deployment remains necessary.