The invisible-grid problem
Most invoices and receipts have no visible gridlines at all — alignment alone tells you where the columns are. Rule-based converters rely on those lines, which is why they shred financial tables into unreadable fragments. The Docling Table & Invoice Extractor uses a tabular vision model that reads the geometry instead: borderless grids, merged header cells, multi-column financial statements, and qty/description/price line items all come back in the right cells.
Pick the output your pipeline needs
Choose Markdown pipe tables for docs and wikis, structured JSON (the document AST) for spreadsheets and automation, Word tables for reports, or tab-delimited text for quick pasting. Everything runs on the same engine as the universal OCR card, so a mixed stack of invoices, statements, and bills can go through one familiar workflow.
Accountants, analysts, and expense auditors: this is the fastest route from a shoebox of PDFs to data you can query.
Sanity-checking the extracted data
AI extraction is fast, but a thirty-second review protects your books: confirm the invoice total matches the sum of the line items, check that tax rows landed in their own cells, and glance at any merged header cells that span columns. Markdown output pastes straight into Notion or Excel, JSON drops into scripts for bulk processing, and Word tables are ideal when the numbers are headed into a formal report. One more shortcut worth knowing: for grid-ruled bank and financial statements, the PDF to Excel card is often the faster sibling — it converts ruled tables directly into .xlsx workbooks with one sheet per table.
Two traps are worth avoiding: multi-page statements where a table continues across pages (extract each page range separately if columns shift mid-document), and stamped or handwritten annotations sitting on top of the grid — the extractor reads the printed table, so route anything handwritten through the general OCR card first if it matters to your records.