The Root Problem: Tesseract & Optical Character Flattening
Legacy OCR engines process pixels in linear horizontal scanning passes. When encountering a balance sheet with an overarching header spanning across three quarterly columns (e.g., 'Q1 2026 Consolidated Results'), standard engines fail to understand column hierarchy.
Instead of attributing child numbers to the parent header, the linear bounding box detector assigns text coordinates indiscriminately, causing subsequent columns to shift to the left.
"A single displaced column in an audited financial ledger turns an automated workflow into an hour of manual spreadsheet repair."
Our Approach: Dual-Stage Computational Geometry & Table Detection
At PDFparse, we developed a two-stage pipeline. The first stage does not read text; it evaluates whitespace distribution and topological line grids to build a mathematical table graph.
Once bounding cells are structurally verified, our localized OCR engine populates each discrete coordinate. This guarantees that numbers cannot 'spill' into neighboring cells even on borderless, scanned documents.
Native Formula Reconstruction
When exporting to Excel (.xlsx), rather than exporting static string numbers, PDFparse analyzes arithmetic relationships. If row 10 is mathematically the sum of rows 1 through 9, we construct native =SUM(A1:A9) formulas in the resulting workbook.
This preserves the dynamic integrity of the spreadsheet for downstream financial modeling.
Key Takeaway
By treating document conversion as a topological graph problem rather than a raw text extraction task, PDFparse achieves 99.8% structural table fidelity across multi-page financial filings.