Grounding: where the box of a value lands

Where a value's box lands on the page: on its own table cell or form field, on its row or line, or only on its whole block. Each row is one way of boxing the same two pages.

Key takeaways

• Mistral OCR 4.1 and Reducto r-1 box blocks only: all 20 values end at their whole table or paragraph.
• Sheaf's Grounding Matching v1, over Mistral's text, boxes 16 of 20 values in their own cell or field, and none in the wrong place.
• Landing AI DPT-3 Pro's own boxes reach 17 cells or fields and 3 lines. These two pages averaged $0.035 a page on its priority tier, against $0.005 for Mistral; its standard tier is half that.

The scoreboard

How it was measured

Two pages of the CFPB's sample loan disclosures, a fictional loan: the cost details of a Closing Disclosure as a digital PDF, and the first page of a Loan Estimate as an image with no text layer. A person checked 20 values on them, each with the place it is printed: 14 table cells, 4 form fields, a choice and a checkbox. Every option read the 150 DPI page image. An option is asked for the smallest box it has for a value, and the box counts when it sits on the value. Sheaf's rows run the product's own code over Mistral OCR 4.1's reading: Grounding Matching v1 matches each cell, field and line to the page's words, from the PDF's text layer or from Tesseract on a page without one; None leaves Mistral's block boxes. The engines' rows use the boxes in their own answers. Twenty values on two pages is a small test: a difference of one or two is noise.

Log