OCR gallery: hard pages, read word by word
No OCR is best on every document. Pick a hard page and see what each pipeline hands back: text, boxes, confidence, and the JSON a call returns. Which means: you see the result before you install anything.
Every answer follows the API reference (schema Result). The pictures are public test pages, never customer data. All output formats at a glance.
Agenda slide with disclaimerLight type on a dark blue ground, plus a long block of small print.
Portfolio slide with hazy photoDisplay type next to a photo with haze, and almost no words to anchor on.
Skewed invoice with stampA tilted scan with noise and a round stamp over the amount.
Balance sheet turned 90 degreesThe whole page lies on its side and has to be turned before it reads.
Equity statement, 12 columnsTwelve narrow number columns with dashes and minus signs.
German table, 40 rowsSix number columns with thousands points, set in small type.
Contents page on red gradientThin light type on a colour gradient, with indents and page numbers.
Double-page balance sheetTwo pages in one image with 11 number columns.
Balance sheet with grid linesSmall values such as 90 sit between ruled lines.
Key figures in two tablesA clean page where the tables carry the content and the layout has to be right.
German invoiceAn invoice table with thousands points and decimal commas.
English invoiceAn invoice table with thousands commas and decimal points.
Balance sheet, two columnsGrid lines, signs and totals that must stay in their row.
Application form with letter boxesSingle letters in boxes and tick boxes instead of running text.
Two-column report textHyphenated words and umlauts running down two columns.
Light text on black pageThe same text as the page before, with the colours reversed.
Thermal shop receiptA narrow strip in typewriter type with many short lines and prices.
Dense quarterly tableEleven number columns and 14 rows in very small type.
Scanned invoice with PAID stampScan noise, a tilt of 1.5 degrees and a stamp across the total.
Scanned valuation modelThirteen forecast columns and two 7 x 7 grids in 5 pt type.
Circus invoice with Paid stampA circus invoice with a printed logo, typed lines, a handwritten date and a Paid stamp across the text.
Hay and flour invoiceA feed merchant's invoice with a table of bags and prices, a handwritten Paid note and thin print.
Rubber goods receipt, 1894A coloured shop receipt with an ornate letterhead, a rubber stamp, handwritten items and a cancel mark.
Telegram with number blocksA faded wire message holding a block of five-digit code groups and a handwritten decode below.
Ship telegram formA government telegram form with typed capitals under a printed form header.
Typewritten diplomatic cableA confidential State Department cable in typewriter capitals with a handwritten circle and file numbers.
Letter received by faxA congressional letter received by fax, with the fax header line, thin type and a ragged right edge.
German delivery note, handwrittenA delivery note of an Innsbruck wholesale grocer with letterhead, items in Kurrent script and a stamp.
Delivery note with stamp and holesAn East German delivery note with carbon copy, punch holes, company stamp and handwritten entries.
Machine factory invoiceAn invoice of a Leipzig machine factory with letterhead, numbered items, a price column and a handwritten total.
Sausage factory receiptA Berlin receipt on brown paper with a stamp, a telephone number and handwritten items.
Furniture dealer receiptA Berlin note slip with one sofa, a down payment and a handwritten receipt.
Almost empty delivery noteA Vienna delivery note with a single handwritten item on an almost empty form.
Orange delivery noteAn orange delivery note with a coat of arms, a triangle mark and handwritten entries.
Carbon copy with strike-outsA typewritten carbon copy about a theatre with struck-out lines and handwritten corrections.
Partly filled payment slipA postal transfer form with a machine-readable line and fields, partly filled in.
The test pages load here. The test pages are synthetic or own pages. The API reference shows the answer format in full.