Table Extractor

It pulls the table out of a document and returns it as structured rows you can download. Point it at any PDF that contains a table — a bank or account statement, a payslip, a tax return, an invoice, a quotation, a report, or a datasheet — and it captures the document title, the table caption, and every row of the dominant table.

As a real table. The rows return as an array with one object per data row, and each row's named fields are the table's own columns — so a bank statement comes back with date, description, debit, credit and balance columns and a payslip with description and amount. Every row and every column is captured and it exports as CSV, XLSX or JSON rows rather than one text blob. This is the point of the rebuilt tool: the old version truncated every table to four fixed columns.

Any document with a tabular section — bank and account statements, payslips and payroll slips, income-tax returns, invoices and quotations, inventory and expense reports, and technical datasheets. It reads both digitally generated PDFs and scanned pages, and column labels are read in several languages (German Tabelle and Zeilen, Spanish tabla and filas, French tableau and lignes).

It uses the table's own column headers as the field names where the document prints them (Date, Description, Amount, Balance, and so on). When a table has no header row the columns are named by inference from the values. The names are auto-inferred by the extraction engine rather than fixed in advance, because the tool is document-type-agnostic — an unusual layout may get a slightly generic label, but no column is dropped.

The tool captures the largest, dominant table on the document — for example the transactions table on a statement rather than the small summary box beside it. For a document whose main content is one big table (a statement, a register, a price list) that is exactly the table you want. Two scalar fields also come back where they are printed: the document title and the table caption.

PDF only, up to 10MB and 100 pages. The document is processed via the Talonic API for extraction only, is not retained for training, and is not shared.