How we measured extraction accuracy
The full write-up is not published yet. This page is a placeholder.
We quote Rapidflare at 95%, Contextual AI at 62%, and AWS Textract at 49% on extraction accuracy. A claim like that only means something if you can see how it was measured, so here is where we are: the definition and the scoring rule are settled, the published detail is not.
What we measured
Extraction accuracy: whether the engine reads a value out of a real technical document correctly, including tables, multi-column layouts, and values that span a drawing and its notes.
How we scored it
Each extracted value is checked against a human-verified ground truth. A value counts as correct only when it matches the source exactly, unit included. Partial reads do not count.
Still to publish
We have not yet published the document set (its size and where the documents came from), the dates the runs were made, or the versions and settings used for each tool in the comparison. Until those are here, treat the three numbers as our own reported figures, not as an audited result.
Want the detail sooner?
Ask us and we will walk you through the run. Better still, bring a document of your own that usually breaks AI, and we will show you what the engine does with it. Get in touch.