Smart Extraction of Structured Data from Unstructured Documents

This thesis presents DataDig, a mobile application designed to address the growing need for efficient and accurate data extraction from documents and images. DataDig uses a combination of Optical Character Recognition (OCR) and Large Language Models (LLMs) to automate the extraction process, leaving the users the ability to define fully customizable templates that retrieve specific information without the need for model training or example documents. The app’s key features include support for various data types, dynamic table extraction, context-aware interpretation and direct integration. Through testing and user feedback, DataDig demonstrates its effectiveness in accurately extracting information from diverse document types. The mobile-first design, coupled with the number of options available for both the process and output format, as well as the added cost transparency, makes it a useful tool for individuals, small businesses, and professionals looking to streamline document processing and data management workflows. This research contributes to the field of data extraction by providing a practical solution that utilizes the power of LLMs to automate this previously time-consuming and error-prone process. The thesis also explores the broader context of today’s AI-powered data extraction, analyzing ethical implications and the current market landscape, comparing DataDig with its main competitors, and discussing potential future developments.