Accuracy Benchmarks

DocSolved OCR

DocSolved's OCR layer reads business documents: printed invoices, receipts, scanned paperwork and camera photos of paper. Camera photos are the hard case: faint text on glossy pages, perspective distortion and local-script diacritics.

• Clean printed pages: in our tests, every word on a printed invoice, a printed receipt and a dense text page was read.
• Camera photos: the current OCR configuration reads markedly more words than our previous one, Polish diacritics included.
• Camera photos remain the weakest case: on the photo we test, the OCR layer read none of the numbers, such as amounts and dates. A value taken from a photo may then have no matched source location, so check it against the document.
• Straightening a photographed page before recognition (camera-document preprocessing) improved photos further in our tests but made already-flat pages worse, so it is never applied to PDF pages.

Structured e-invoice fields skip OCR entirely; they come from the source data.

DocSolved AI Algorithms

Beyond OCR, DocSolved applies document-intelligence algorithms that turn raw text into structured, auditable fields.

• Multi-invoice splitting. When a single PDF carries several invoices back-to-back, DocSolved detects the page boundaries from the embedded text layer before any OCR or model call, and offers them to the user for review rather than applying them silently. On a synthetic test set with known boundaries it found every boundary in clean text and got a few wrong once OCR-style noise was added, which is why it suggests rather than splits.
• Field extraction. Values are returned with the scope the matcher could establish: an extracted field may carry a confidence score, matched source text, a match status, and a page location when the value is anchored.
• Suggest, then confirm. Split boundaries and follow-up actions are offered as suggestions, and take effect only when a person confirms them.

How DocSolved compares

The measurements on this page compare DocSolved with its own earlier configurations. We have not published a head-to-head comparison with other OCR or document-processing products.

The most reliable comparison is on your own documents: the live demos need no account, and you can start free without a card. For help setting up a side-by-side test, write to [email protected].

Methodology Notes

OCR is scored by how many words of a known text it reads (recall) and how many of the words it outputs are right (precision): against a human transcription for a camera photo of a real Polish document, and against the source text for sample printed pages in English. Split detection is scored on a synthetic set of multi-invoice PDFs with known boundaries.

The samples are small, one document per OCR case, so treat the results as indicative rather than as a promise for your documents. No third party has reviewed them. They come from committed internal evaluation artifacts; this page summarises them rather than republishing the raw figures.

Not measured yet: field-level extraction accuracy (whether the right value lands in the right field), end-to-end processing time, and OCR accuracy for languages other than Polish and English.

Questions about the methodology or results: [email protected]