# Audion Office OCR AI

[Русский](README_RU.md) · [User Guide](USER_GUIDE_EN.md) · [History](CHANGELOG_EN.md)

**Contents**

- [Why It Exists](#why-it-exists)
- [What Is in the Package](#what-is-in-the-package)
- [What Is Assembled From It](#what-is-assembled-from-it)
- [Principles](#principles)
- [Next](#next)
- [Technical Reference](#technical-reference)
  - [Source and Target](#source-and-target)
  - [Second Pass](#second-pass)
  - [Full Archive](#full-archive)

Recognising scans and converting office documents: from paper to an editable
document, a spreadsheet, a searchable PDF, and everything else — from a single
recognition pass.

## Why It Exists

Recognition costs money and time. The usual workflow ignores that: you run a
scan, get a Word file — and a week later you need the spreadsheet. So you run it
again. Then you need a PDF with a text layer — once more.

Every pass means new money, a new queue, and a new result that may differ from
the previous one.

Here the order is different:

```
sources → recognition → document package → review → any format
```

**You pay for recognition once.** The result is stored in an independent
`<name>.document` package, and every format afterwards is assembled from it
locally — with no call to a paid service.

## What Is in the Package

Not the text, but everything the text came from:

* an exact copy of the original;
* images of every page before and after preprocessing;
* hashes;
* the text and coordinates of **all** recognition candidates, not just the chosen
  ones;
* reading order, tables, merged cells;
* confidence for every fragment;
* cross-checks between different engines;
* the provider's raw response;
* room for manual corrections.

**Completeness here means no loss of source data**, not a promise of flawless
recognition. A recognition error is visible and fixable precisely because
everything it was based on lies beside it.

## What Is Assembled From It

Eight results, independently and at once: DOCX, XLSX, PDF with a text layer, ODT,
Markdown, recognition JSON, HTML, and a review sheet.

By default only DOCX is produced — there are no format presets, nothing
superfluous is generated. Reassembly runs locally; no third-party office suite is
required. The maximum page size is A3.

## Principles

**The second pass by another engine is independent.** It receives the same
cleaned raster but does not see the first engine's text. So a disagreement
between them means genuine ambiguity, rather than a hint repeated back.

**Tables are parsed by a physical grid.** For ruled tables a grid is built from
the long lines of the raster, and the column count is verified against a `1…N`
numbering row taken from the coordinates of recognised words. Only then are the
engines compared inside exact cells.

Merges are restored **only in the header** — there they are predictable. In the
body of a table a merged cell more often means a parsing error than the author's
intent.

**Disagreements over digits are preserved.** If the engines read a number
differently, both readings stay in the package. A number is where a recognition
error costs most, and a person should settle it.

**Markdown is not the centre of the system.** It is convenient as a control
format and as input for language models, but the document package is not built
around it.

PDF files are read and written by the PyMuPDF library, which is distributed under
the GNU AGPL 3.0. The program's own code stays under Apache 2.0, but the build as
a whole — the program together with PyMuPDF — is offered under the terms of the
GNU AGPL 3.0; see the `NOTICE` file.

## Next

* [User Guide](USER_GUIDE_EN.md) — step by step.
* [History](CHANGELOG_EN.md) — what changed.

---

## Technical Reference

### Source and Target

The path rows are called **Source** and **Target**. The source may be a single
file or a folder - picked, typed, or dropped on the window; the history of paths
keeps pinned paths on top. **Reset** sets both back to the folders of the program,
**List** names what lies in the source, the bin clears the source, the target or
both into the Recycle Bin.

An external source is copied into the managed `input` folder before a run, and
the results are copied into the chosen target. The window asks first when `input`
or `output` hold anything - they are cleared before the copy.

### Second Pass

| setting | what it does |
|---|---|
| none | the main engine only |
| Tesseract | local cross-check |
| Yandex | cloud cross-check |
| Yandex and Tesseract | both |

### Full Archive

Packing the whole package into one archive remains an engine capability for
compatibility, but it is not shown among the main buttons: it is rarely needed and
takes a great deal of space.