Extract Tables

PDF to Excel returns a single XLSX with every line of the PDF's text in one sheet, each value placed in the column its position puts it in, so it suits getting all of the text into a grid. Extract Tables looks only for table-shaped blocks, leaves out the titles, captions and prose around them, and gives each table it finds its own sheet and CSV file. Both keep 1,250.00 whole and leave an empty cell empty.

The pages with text are read as usual, and the scanned pages simply produce nothing. A file with no text on any page stops with a message telling you to run OCR PDF first; add a text layer there, then bring the searchable PDF back.

Columns are found from the white space that runs down the rows. When a long cell in some row fills the gap between two columns, there is no gap left and those two columns are read as one for that table. A heading spanning two columns over the rows beneath it is allowed for. Two tables side by side on the same rows also read as one wide table. Multi-word cells such as North America or 12 March 2024 stay in one cell, right-aligned amounts stay in one column, and merged header cells are not reconstructed.

A block counts as a table only when at least two columns each hold something in two or more rows, one column holds short values of fewer than 4 words each (and is not mostly bullet items), and no sentence wraps from one row into the next. Two columns of prose, side-by-side bulleted lists and letterhead lines fail those tests on purpose. A table in which every column holds longer text can fail them too, as can one cut into single rows by one-line subheadings. If nothing in the file passes, the job ends with the message that no table-like structure was detected.

Each page is analysed on its own, so a table spanning three pages comes out as separate CSV files and sheets, at least one for each page whose part of it is detected as a table, and a header row the PDF repeats on each page is repeated in each sheet too. Stack the sheets and delete the repeats to rebuild it.

Yes, when a cell holds exactly one number. Values such as 1,250.00, $4.20, 15% or -4 are stored in tables.xlsx as numbers with a matching format, so they can be summed straight away. A code with a leading zero such as 00123, a number of more than 15 significant digits, a European 1.250,00, a date or a bracketed negative stays text exactly as printed. The CSV files have no byte-order mark, so if accented characters look wrong after a double-click, open them through Excel's text import.

Up to 100 MB per file.

A ZIP with one CSV file per table it found (table-1.csv, table-2.csv and so on) and tables.xlsx, an Excel workbook holding every table on its own sheet, named after the table's number and page.

By where the text sits: words set a normal space apart are joined into one cell, columns come from the white-space gutters that run down the rows, and a block is kept only where at least two columns hold values in two or more rows and one column holds short values rather than sentences. Ruled lines and cell borders are not read, so a table is found by its alignment alone.

No. A scanned page is a picture with no selectable text, so the job stops and says so rather than guessing. Run the scan through OCR PDF first, then extract from the result.