v1.0.11.0 is here: six tools, one workspace.See what's new →
Armor EdgeAI
EN
Download

Module 03 / 06

Document Recognition

Scans, PDFs, presentations and spreadsheets get layout analysis and text recognition on-device; the result can then become a report, or be lined up segment by segment against another version.

The Document Recognition workstation: document list, page preview and recognition results

FEATURE 01

OCR analysis

Two modes: advanced does full layout analysis with high-accuracy recognition; light hands the whole page to a vision model for fast plain text.

  • Advanced: structure preservedHeadings, paragraphs, tables and formulas each become their own block, and the output can be reviewed and edited block by block — suited to reports and contracts.
  • Light: fast textThe whole page goes straight to plain text, which suits screenshots and receipts; no layout analysis, and no OCR-specific model to download.
  • Several export formatsMarkdown, Word, HTML, JSON, plain text, CSV.
  • More than scansPowerPoint and Excel files are read directly; recognize every document in the list at once, or have recognition start as soon as you upload.
Document Recognition — OCR
Heading
Paragraph
Table
Caption
Analyzing layout
Output
# Q3 operations summary

Revenue grew over last quarter, driven mainly by…

| Item    | Q2 | Q3 |
|---------|----|----|
| Revenue | …  | …  |

*Figure 1: Monthly shipments*
Layout and table structure preserved

FEATURE 02

Insight reports

A recognized document can be worked up into a report with several angles on it — not just a paragraph of summary.

  • A dozen or so sectionsKey summary, keywords, chapter outline, action items, notable excerpts, mind map, glossary, open questions, and more.
  • Jump back to the sourceChapters and excerpts are clickable and take you straight to the matching page in the document.
  • Ask and exportQuestion the report in the drawer on the right, or export it as HTML or Markdown.
Document Recognition — InsightsGenerating
Key summaryDone
KeywordsDone
Chapter outlineDone
Action itemsGenerating
Glossary

FEATURE 03

Document compare

Take two recognized documents: the pages are aligned by algorithm first, then the changes are worked out segment by segment.

  • Align first, compare secondPage correspondence is decided by an algorithm; only the pages judged to have changed go to the language model for segment-level analysis.
  • Three viewsVersion one only, side by side, or version two only — plus a toggle for showing just the segments that differ.
  • Honest flagsPages that couldn't be matched confidently are marked "low pairing confidence, please verify manually" — no pretending it's 100% right.
Document Recognition — Compare4 changes
Version 1Version 2
This section describes the overall architecture of offline inference.This section describes the overall architecture of offline inference.
Latency depends on hardware.Latency depends on local hardware, not datacenter load.
—So there is no queueing or slowdown at peak hours.
The cloud fallback process will be evaluated separately.—
All models are stored on the device.All model weights are stored on the user's device.
AddedRemovedModifiedUnchanged

Color isn't the point — the marker and weight on the left edge are the change type

Next

Got the machine? Four steps to start

Install, agree, activate, download models — most of the time is just the download.