AudionAudion

Audion DocFlow (Portable)

USER_GUIDE_EN.md

Audion DocFlow (Portable)

Contents

A portable, offline-friendly toolkit for deterministic document cleanup, compliance gating, and table helpers around DOCX, XLSX, CSV, and Markdown.

The project is designed for a stable portable workflow:

The current main-kit scope is not OCR. OCR experiments and AI-assisted OCR flows should live in a separate dedicated project instead of being reintroduced into this deterministic office-helper kit.

What it does

DOCX

report\ plus per-document .md/.json files in a same-named report folder, with file names matching the source documents

with human-oriented locations (page / section / Table N / Figure N). The implemented RO layer checks tables, table/figure captions, sections/orientation, headers/footers and page numbering, fields/TOC/links/ bookmarks, lists/numbering, duplicated blanks, and dangling headings; the safe correction command writes copied DOCX files under output\docx_anomaly_fixed and can collapse duplicated blanks, remove exact table row heights, enable table-cell text wrapping, and optionally normalize table borders/cell margins

text hygiene and deterministic audit rules run first, and document anomalies can be attached as the second pass. The unified docx_deep_hygiene.md/.docx/.json report records pass records, domain_boundaries, ownership boundaries, target_index, unified_findings, target/issue fingerprints, and suppressed exact duplicates. Morphology-backed address/toponym validation remains a future separate layer.

existing DOCX without rebuilding document text; it infers the body font, uses two density-based table font sizes, normalizes borders and cell margins, skips the first two document pages by default, balances columns, and fits widths to document sections; the default cell margin is 0.2 cm

rows preserved as merged rows; the Unifier preserves detected multi-row and merged headers more carefully, with pre-header rows, A4/A3, orientation, and millimeter margin controls

tables to current document margins / explicit page setup without rebuilding the whole document; small-column tables can be excluded from fitting by a default threshold of 3 columns

available text indent because XLSX has no true Word-like cell-margin property

XLSX, CSV, and Markdown

GUI

Source and the chosen result folder as Target directly to the backend, without staging a copy in local input\

config\gui_settings.yaml, while palettes and CSS tokens live in config\ui_colors.yaml; the default is Code Dark (code_dark)

morphological replacement automation

and single-file or batch anchor removal; the RULES button opens config\rules\

ANOMALIES, and DOCUMENT STYLES` tabs; scan/fix modes use distinct green/orange outlines, checkbox classes are laid out in a responsive grid, and the run button stays in the top command row beside Back

export, dedicated repeated-header reconstruction, the root Unify tables in document command, safe and width-only unification, width optimization, orientation adaptation, table fitting, and table extraction

description stays separate; parameterized commands keep lighter rounded blocks with dark fields and checkbox/radio chips

fields, so longer explanations do not need to stay visible in the panel

Markdown table export and compact table-cell margins live in Technical operations

and are not shown on the first GUI screen

Launchers

User launchers

Both launchers use the same internal logic, with:

Find and replace, DOCX control and cleanup, Compare and reconcile, WORD/EXCEL TABLES, Technical operations, and Checks;

Service layer

The service layer now relies on:

  1. Put source files into input\.
  2. Run launcher_project.cmd or launcher_project_ru.cmd.
  3. Choose the needed tool.
  4. Read processed files in output\ and reports in report\.

For environment or release tasks:

For the GUI:

launcher_gui.cmd

Reports and UI language

while JSON reports are the machine contract for an external LLM pipeline. This project does not call an LLM; it only deterministically collects candidates, context, and simulated edit results.

UTF-8 JSON and readable Russian strings.

DOCX A/B, Word style ids such as a7) are intentionally kept technical.

the related operation to regenerate a Russian report.

Smoke Tests

Quick smoke:

& '.\runtime\python.exe' '.\tests\smoke.py' --quick

Full smoke for the scripted tool set:

& '.\runtime\python.exe' '.\tests\smoke.py' --full

Project layout

Notes

installed Microsoft Word and available PowerShell (system_core\powershell\pwsh.exe, pwsh.exe, or Windows PowerShell).

project scripts should not be stored there.

Troubleshooting

If the runtime is missing, use:

builder_main.cmd

or:

install\Build_Portable_Env_Build.cmd

If launcher behavior looks wrong:

with install\Check-CmdEncoding.cmd;

License

See the project license file if one is provided, and licenses\THIRD_PARTY_NOTICES.md in release-oriented builds.

Canonical Workbench labels

Workbench uses the same Audion Image Tools public vocabulary in every project. Its buttons always keep the same order and labels: Source, Add file..., Target, Reset, Delete, List.

Reset returns to project input/output and does not delete files; Delete clears the current Source and Target only after confirmation. The exact Russian labels are Источник, Добавить файл..., Назначение, Сбросить, Удалить, Список. The Workbench variants Destination, Clear, Цель, and Очистить are not used.

Production Workflow

Start with a small sample that contains the document structures expected in the full set. Select a separate output directory, run the operation, review the report, and open every sample result. Only then repeat the same settings against the full folder.

For exact replacements, verify the requested text, scope, case behavior, punctuation, and whether matches may cross run boundaries. For morphological replacements, inspect the generated candidate list and reject contextually incorrect forms before producing final documents.

For deep DOCX hygiene or anomaly repair, compare styles, numbering, tables, headers/footers, fields, tracked changes, and section breaks. The absence of a parsing exception does not prove that Word layout is unchanged.

Folder Runs

Recursive runs must exclude the selected output, runtime, work, workspace, reports, caches, and prior generated copies. Review the discovered file count before execution. Use stable names or a result manifest so that every output can be traced to one source.

Reports And Recovery

The report distinguishes changed, unchanged, skipped, and failed files. Keep a failed file and its diagnostic entry together. Correct the operation or source and rerun only that file or a controlled subset instead of repeating already accepted documents.

Final Review

Compare input/output counts and filenames, open representative documents in Microsoft Office or the intended consumer, and inspect areas affected by the operation. Preserve sources and reports until acceptance. Cleanup is performed only after useful output has been moved out of disposable work folders.

Treat the GUI manifest as the authoritative catalogue of selectable actions and controls. The guide adds the human layer: select an operation by the result you need, read its tooltip, confirm the source and target, and use the report to prove what the backend actually changed. When a control appears conditionally, document the condition and its consequence instead of assuming every operator will infer it from the screen.

A release review should therefore compare the manifest, GUI labels, command preview, report fields, and this guide. Any mismatch is a documentation or interface defect even if the underlying conversion succeeds.

Alle Projekte