Το work with title Optimization of enterprise workflows through automated information extraction from PDF files using large language models by Athanasakis Evangelos is licensed under Creative Commons Attribution 4.0 International
Bibliographic Citation
Evangelos Athanasakis, "Optimization of enterprise workflows through automated information extraction from PDF files using large language models", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2025
https://doi.org/10.26233/heallink.tuc.105015
The large volumes of files circulated in today’s enterprise workflows have promptedthe development of methods based on Artificial Intelligence (AI) techniques for automated information extraction, retrieval, and summarization. In this diploma thesis, methods for processing and extracting data from semi-structured Portable Document Format (PDF) documents are studied and implemented using Large Language Models (LLMs). The project is divided into two distinct parts. In the first part, the study focuses on information retrieval from Greek soil analyses, which are characterized by their heterogeneous structure and formatting. Various text extraction techniques are examined, both from natively digital and scanned documents, using Optical Character Recognition (OCR). The contribution of individual sub-modules in the processing pipeline, such as post-processing for text extraction error correction and translation from Greek to English, is then investigated to the accuracy and efficiency of the overall structure. Various information retrieval techniques are then compared, including the full-context prompting approach and Retrieval-Augmented Generation (RAG), with the goal of evaluating the efficiency of each processing flow. In the second part, the methodology is generalized to be applicable to PDF documents from any domain. To this end, three agents are developed: The Field Detection Agent identifies candidate fields, the Post- Processing Agent filters and normalizes the results, and the Prompt Builder Agent dynamically constructs prompts for the information retrieval phase. Different architectures created by these agents are examined for extracting the names of fields that can be retrieved from the document. The efficiency and accuracy of the best information retrieval method from the first part is then re-evaluated, along with a variation of the full-context prompting approach. The proposed approach allows for automatic, adaptive, and efficient information extraction from a variety of texts. Overall, the thesis contributes to both the evaluation and improvement of different processing flows for data extraction from Greek soil analyses and the development of a general and scalable multi-agent architecture for any domain. The proposed framework can be applied to various fields, enhancing the automation and accuracy of information extraction fromPDF files.