Structural: An Intelligent Visual Document Data Extraction and Reconstruction

Abstract

Financial professionals routinely process large volumes of paper-based documents such as receipts, invoices, purchase orders, and utility bills. Because these records typically reach accounting systems as scanned images or mobile photographs, their data cannot be used directly by digital workflows. This paper presents Structural, a browser-native intelligent web application that automates the extraction, structuring, and multi-format export of financial data from document images. The system performs Optical Character Recognition (OCR) entirely client-side using Tesseract.js, a WebAssembly port of the open-source Tesseract engine, eliminating the need for a server. Recognized text is processed through a five-stage rule-based Natural Language Processing (NLP) pipeline that performs document-type detection, regexbased metadata labeling, line-item parsing, summary-field extraction, and supplementary key-value capture. Structured output is exported in CSV, JSON, and Microsoft Excel formats via the SheetJS library. Because all computation occurs on the client device, the system requires no API key, subscription, or backend infrastructure, and user data never leaves the browser. On a 50-document evaluation set, Structural achieved a Field Extraction Rate (FER) of 88.5%, a Line Item Accuracy (LIA) of 84.2%, a Character Error Rate (CER) of 6.1%, a Word Error Rate (WER) of 9.8%, and 100% export correctness across all three output formats. Relative to transformer-based document-understanding models, Structural trades a measurable accuracy gap for zero infrastructure cost, complete data privacy, and offline capability — a trade-off this paper positions explicitly rather than treats as incidental.