This thesis builds and validates the data infrastructure to solve this problem. Following a Design Science Research (DSR) methodology, the system was developed as a robust Proof of Concept (POC) to address specific resource and privacy constraints. An initial attempt using YAML failed to produce valid files, leading to a successful pivot to a JSON-based templating approach. The core innovation is a custom recursive XML parser that penetrates nested tags in complex legacy documents. In testing against a representative corpus, this parser increased table detection from 4 to 243—a 60.75× improvement over standard libraries.The infrastructure is now validated as production-ready. It captures 92.19% of semantic content and maintains 86.3% recall for technical equipment tags. It is also fast: search latency stays under 20ms, and the overall processing time drops from 6-9 hours of manual work to just three seconds.

In conclusion, this work proves that by solving the parsing problem and enforcing a JSON “Single Source of Truth,” messy legacy documents can be transformed into highquality training data. This establishes the necessary validated foundation for future full-scale model training and diagram integration.