Analysis is the layer everything else depends on. DocxIntel resolves page structure, reading order, tables, stamps and handwriting across Arabic and English — running entirely on your hardware, with coordinates and confidence attached to every block it returns.
Accuracy is measured field-by-field against a human-adjudicated ground truth, not character-by-character. The methodology, document mix and per-field breakdown are published in full on the accuracy page.
Most extraction errors are not extraction errors. They are structure errors that happened one step earlier — a merged table, a mis-ordered column, a stamp read as a character.
Headers, body text, columns, footnotes, captions and marginalia are separated and ordered the way a human reads them, including right-to-left flow on Arabic pages.
Cell boundaries, merged cells, spanning headers and continuation rows across page breaks are reconstructed into an addressable grid.
Handwritten field entries, marginal notes, ticked checkboxes and initials are recognised as content rather than discarded as noise.
Overlapping ink is separated from the printed text beneath it, so a stamp across an invoice total does not corrupt the total.
Skew, warp, glare, shadow, low contrast and generational fax loss are corrected before recognition rather than passed through to it.
Each block, cell and token carries its page number, bounding box and confidence score, so any downstream value can be traced back to the pixels it came from.
Analysis output is not a flat blob of text. Every element keeps its position on the page, which is what lets a reviewer, an internal auditor or a regulator see exactly where a number came from.
This is the difference between a tool that gives you data and a tool that survives an audit.
What the analysis service accepts and returns in a standard deployment.
Format coverage in a specific deployment is confirmed in writing during the Proof of Value against your actual document mix, not against a generic list.
Cloud parsing APIs are good software. They are also priced per page, and every page has to reach them to be read.
Where pages are processed
True air-gapped operation
Cost of re-analysing 2M archived pages
Model weights
Self-hosting available on
Arabic as a primary script
Comparison reflects publicly documented behaviour of metered document-parsing APIs as of 2026, including credit-based per-page pricing and enterprise-gated self-hosting. Verify against current vendor documentation before making a decision — we would rather you checked.
Watched folders, S3-compatible object storage, your DMS, an email intake mailbox, or a direct API call. Nothing is uploaded anywhere — the service reads from the storage you already control.
Deskew, dewarp, denoise and contrast correction run first, so recognition sees the best available version of each page rather than the version that came out of the scanner.
Layout segmentation separates regions and establishes reading order, including right-to-left flow, before any text is transcribed. Tables are reconstructed as grids, not as loose text.
Printed text, handwriting, stamps, seals and checkbox states are recognised by models running on your GPUs, each result carrying coordinates and a confidence score.
Structured JSON, markdown, annotated PDFs or tabular exports are written back to your storage or returned over your API, ready for identification, classification and routing.
The technical objections that come up in every evaluation, answered without hedging.
No. The layout, OCR and table-structure models ship inside the deployment bundle and load from local storage. A fully air-gapped install performs identical analysis to a connected one — there is no fallback call to an external service, so there is nothing to firewall.
