Tensionix
Audion Office OCR AI

v2.0.0 ·

Changelog

EN

Changelog

Contents

2026-10-15 - 2.0.0: a native window and a console command

Added

  • A native Windows window (C#, Avalonia): Start.exe opens App\AudionOfficeOcrAi.exe. The window carries its own .NET and its own PowerShell (App\AudionPs.exe), so nothing is installed.
  • SETUP in the header, before SETTINGS: the installation of what the machine needs for local OCR; while the machine has no Tesseract, the window opens on it at start. Five sections in one band under the header: LOCAL OCR, API OCR, OCR CHECK, TABLES, TOOLS; the quick conversion of office files into Markdown is TOOLS -> To Markdown. The engines of LOCAL OCR and the providers of API OCR are switches over the form; a form stands in named blocks with a coloured stripe, rarely needed settings under ADDITIONAL, risky choices in red. The marks of Yandex, Mistral, Gemini and ChatGPT wear their brand colours.
  • ocr_cli.cmd: every operation of the window runs from a console with the same settings - describe, run --operation ID [--source] [--target] [--set NAME=VALUE], options (--refresh asks the provider), key add|remove, hardware, version. Exit codes: 0 done, 1 failed, 130 stopped, 2 a wrong call - refused before anything runs. The window writes the line it runs into its terminal.
  • The header names what the machine has for OCR: the recommendation for its video card, Tesseract and its languages, the layout service of Surya, the providers that have a key - a count of key files, never a key.
  • The layout service of Surya is started and stopped from the line under the SURYA switch.
  • A key of a provider is added and set aside from its field: it is pasted into a dialog, goes to the console command by its standard input and is shown nowhere; a key set aside moves to config\deleted_keys\.
  • The lists of models open under their field at once, from what the program keeps; REFRESH THE LIST asks the provider.
  • The working area: SOURCE (a folder or one file, dropped on the window or picked) and TARGET with a history of paths and pins; LIST names what lies in the source.

Changed

  • A run with a SOURCE outside the program asks first when input\ or output\ hold anything, and names how many items each holds: the source is copied into input\, and both are cleared before it. A SOURCE inside output\ and a folder inside input\ are refused before anything runs.
  • One clearing for the working area: the contents of SOURCE, TARGET or both into the Recycle Bin, and the DocumentModel artifacts, each with its question; the root of a drive, the program itself and links are refused.
  • STOP asks the operation to stop at its next safe point and ends it at once on a second press; closing the window while an operation goes asks first and waits for the stop.
  • The terminal shows the colours of the operations and one live line of what a run is doing.
  • The output of the programs an operation runs reaches the terminal as text whatever they write in: UTF-8, the OEM and the ANSI code pages, and now UTF-16 too; the byte order mark of UTF-8 no longer stands at the start of a line. The lines of the Surya layout service script, Russian messages of PowerShell among them, are no longer question marks.
  • Layout.Service.ps1 -Disable stops only the layout service of its own copy of the program.
  • The environment is built anew from the current packages of install\requirements_full.in; with google-genai 2.31 the request to Gemini names the type of a page image mime_type.
  • The install scripts take App\AudionPs.exe first. install\verify_portable_env.cmd checks the console command, Tesseract, Start.exe and the self-check of the window. builder_main.cmd gains START LAUNCHER and WINDOW; the steps of the OCR engines keep their numbers.

Removed

  • The window on NiceGUI in a pywebview frame, the console menus (launcher_gui.cmd, launcher_project.cmd, launcher_project_ru.cmd, launcher_tools.cmd), fzf and the portable PowerShell. The licences are collected by Audion Build Licenses and the release is made by Audion Build Publisher, so install\make_release_archive.cmd and system_core\license\ are gone too.
  • NiceGUI and pywebview leave install\requirements_full.in; the environment built next holds only what the engine needs.
  • What nothing in the program used: the HTML terminal of the former window, the decoders and helpers only it called, a stub of Yandex batch OCR, the second example folder of Yandex keys under system_core\ocr_brick\config, the unused requirements.txt of OCR Brick and ocr_named_entities_ru.md (the hints are read from ocr_named_entities_ru.yaml).
  • DEV Markdown PDF with its engine and its Chromium (runtime\.playwright, step [05] of builder_main.cmd, the package playwright): paired dark and light PDFs of Markdown are made by Audion Build PDF Render (render_markdown.cmd). The PDF of the program stays Word-based (Build PDF).

2026-07-19 - Resilient bootstrap downloads

Changed

  • Restored the Workbench path/deletion helper chain so an unpinned external Source selected from history reaches the operation snapshot exactly like the native picker route.
  • FZF and portable PowerShell installers now resolve and download deterministic release assets without consuming GitHub REST quota in the normal path, while preserving API metadata as a fallback when direct asset conventions change.
  • Real-ESRGAN remains on its metadata-driven release resolver because its Windows portable asset does not follow the same simple latest-version contract as the bootstrap utilities.

2026-07-17 - Canonical Workbench and shared CLI/GUI services

Changed

  • Transplanted the byte-identical canonical Workbench with the shared Source, Add file..., Target, Reset, Delete, and List vocabulary.
  • Added protected single-file/folder pickers, process-tree cleanup, local-only GUI host enforcement and the sanitized native-window shutdown lifecycle.
  • Routed GUI Workbench paths as per-operation Source/Target snapshots while preserving managed input/output roots for the backend mirror contract.
  • Added system_core/cli_operation.py; both CLI/FZF launchers now execute the same manifest/service operations as NiceGUI for OCR Brick, Workbench resolvers and office builders.
  • Replaced obsolete CLI-only OpenAI/Gemini OCR copies with the current Tesseract, Surya, Yandex, Mistral, xAI, Gemini and OpenAI OCR Brick routes.
  • Hardened cleanup and release staging: .git is preserved, while user documents, reports, caches, local route history and Yandex credential metadata are excluded from portable archives.

Verified

  • Added single-file mirror/Target synchronization tests, CLI/GUI dispatch parity checks and an offline Tesseract-to-DocumentModel/DOCX/JSON/Markdown integration smoke.
  • Removed the unused legacy extract_ai service and generic model selector after confirming that no manifest, GUI, CLI or test referenced them.

2026-07-15 - Provider-neutral DocumentModel and office exports

Added

  • Added the lossless <name>.document package as the canonical OCR layer, retaining exact sources/page rasters, hashes, all OCR candidates, coordinates, structured regions, table spans, confidence, verification and manual-correction slots.
  • Added local DOCX, Searchable PDF, XLSX, ODT, Markdown, HTML, OCR JSON and full ZIP exporters. They do not rerun paid OCR and do not require LibreOffice.
  • Added independent checkbox-chip output selection backed by the canonical DocumentModel.
  • Added automatic local Tesseract coordinate generation when a cloud OCR result is exported as Searchable PDF.
  • Capped adaptive DOCX page size at A3.
  • Added content-aware column balancing for DOCX and XLSX: compact numeric/year columns yield space to names, addresses and other long text.
  • Added Mistral second-pass GUI controls: none, Tesseract, Yandex, or Yandex plus Tesseract, with suspicious/all-page scope and optional text fusion.

Changed

  • Output formats now use a balanced 4-by-2 checkbox-chip grid: four office formats and four developer formats. Full archive ZIP remains supported by the backend but is no longer exposed in the primary OCR GUI.
  • Removed the duplicated Output formats and Scan cleanup field labels while retaining their shaded explanatory hint blocks; the section headings Deliverables and Preprocessing remain the only upper headings.
  • Restored visual separation for deliverable chips by placing the complete 4-by-2 checkbox grid on the same subdued dark container treatment used by segmented button groups.
  • Expanded the two OCR output-contract buttons into equal full-row columns for better visual balance.
  • Added conservative Mistral/Yandex fusion. Table cells use verified physical geometry, provider-native coordinates or unambiguous native Tesseract word anchors; unverified grids never authorize replacement. Numeric disagreements and ambiguous coordinates remain review items.
  • Added a physical ruled-table model: long raster lines define rows/columns, the OCR numbering row verifies the column count, header merges are recovered from missing separators, and all provider text is assigned inside exact cells.
  • Added Yandex candidates, provider verification and per-item fusion decisions to DocumentModel 1.1 and the XLSX audit sheet.
  • Fixed the paid OCR root filter so Mistral OCR 4 is actually visible in the GUI.
  • Kept the OCR engine selector in one horizontal row with a single Engine/Движок heading. Engine buttons now divide the full section width evenly: two local engines use two equal columns, and five API engines use five equal columns.
  • Promoted the shared Deliverables/Готовые файлы block directly below the engine selector for both local and paid OCR. Preset buttons were removed; formats are independent checkboxes and DOCX is the working default. The internal DocumentModel is always preserved.
  • Added a confirmed maintenance action Clear DocumentModel artifacts / Очистить DocumentModel. It removes regenerable *.document, *.document.json and *.verification.json artifacts from managed output/report/workspace while preserving DOCX, XLSX, Markdown, PDF, images and sources.
  • Removed the obsolete Build office output card from the main GUI. The old Markdown compilers remain available under Project tools -> Rebuild legacy Markdown for compatibility with archived Markdown workflows.
  • Reworked OCR deliverables into a balanced 4-by-2 grid of rounded rectangular checkbox chips: four equal amber office columns and four equal sea-green dev/machine columns. The chip outline is intentionally very subtle, while the square checkbox keeps a clearer but slightly muted outline; narrow layouts fall back to two columns.
  • Flattened the Tables window into three expanded operation blocks with their own parameters and Run buttons; coordinate extraction, Markdown table inspection and XLSX build no longer require child navigation.
  • Refreshed the Russian and English documentation stack around the current DocumentModel-first workflow, direct 4-by-2 deliverables grid, conservative table geometry, maintenance semantics and legacy-only Markdown rebuild path.

Clarified

  • “100% completeness” means no loss of source data; it is not a claim of 100% OCR accuracy.

2026-07-06 - DEV Markdown PDF and GUI polish

Changed

  • Changed the DEV Markdown PDF default output to docs/PDF in the GUI, service fallback and standalone renderer.
  • Reworked DEV Markdown PDF source mode into compact top-level buttons: Источник, input/output, Документация, PDF-пары, Устаревшие PDF.
  • Kept PDF рядом с MD as an explicit alternative: every folder with Markdown gets one shared PDF subfolder next to those Markdown files.
  • Clarified the other DEV Markdown PDF output modes: Прямо рядом с MD, Общая docs/PDF, and Зеркало в Назначение through the current Workbench Target.
  • Replaced the old external dev-PDF folder wording in the GUI-facing documentation model.
  • Compact hardware recommendation badge is now shown only in Инструменты проекта and Локальный OCR.
  • Kept canonical tooltip styling and detailed control tooltips while skipping noisy per-checkbox tooltips for document-kind chips.

2026-07-01 - OCR Brick integration and UI cleanup

Added

  • Added root-level OCR windows:
    • Офисный текст без OCR
    • Локальный OCR
    • Платный API OCR
    • Тест качества OCR
    • Проверка реквизитов
    • Таблицы
  • Added OCR Brick pipeline under system_core/ocr_brick with rasterize, preprocess, cache, OCR engine dispatch and Markdown/JSON output.
  • Added CPU preprocessing fallback with PIL/NumPy: JPEG denoise, autocontrast, contrast boost, unsharp mask, adaptive binarization, deskew and vertical-line removal.
  • Added Real-ESRGAN ncnn-vulkan portable installer and detection from tools/realesrgan-ncnn-vulkan.
  • Added engine-aware scan cleanup profile Очистка скана:
    • Авто uses tuned defaults for the selected OCR engine.
    • Без очистки sends the page raw.
    • Тяжёлый скан uses stronger denoise, contrast and sharpening.
    • Цифры boosts legal/financial identifiers.
    • Ручной exposes the old detailed preprocessing switches in Advanced.
  • Added local OCR engines:
    • Tesseract as the recommended local baseline.
    • Surya as optional slow/local OCR for cases where API cost matters more than time.
  • Added paid/API OCR engines:
    • Yandex Vision OCR with selectable mode/model and OCR-safe model cache.
    • xAI vision OCR with model list refresh, curated pinned models and API smoke test.
    • Gemini and OpenAI vision OCR through the same OCR Brick vision contract.
  • Added Gemini API cost/scheduling controls:
    • Gemini tier selector with Standard and Flex for regular OCR and one-page OCR quality tests.
    • Gemini Batch OCR project tool to prepare local JSONL manifests and optionally submit asynchronous Batch API jobs.
    • Забрать Gemini Batch project tool to poll/download finished Batch OCR results into Markdown.
  • Added 2-pass Tesseract in paid OCR:
    • For Yandex it writes a local sidecar with Tesseract text/boxes.
    • For xAI/Gemini/OpenAI it can inject Tesseract text as a fallible prompt hint.
  • Added one-page OCR quality benchmark for raw/clean comparisons before large OCR runs.
  • Added numeric/legal requisites checker Проверка реквизитов with Tesseract, Surya or vision-page locator modes and xAI verification.
  • Added global tooltips for command buttons, fields, model selectors, service buttons and API-key controls.
  • Added API key add/delete UI with warnings and provider-specific key selectors.
  • Added pinned model storage in config/gui_model_pins.json.
  • Added hardware recommendation badge for local OCR and project tools. It detects GPU class and recommends Vulkan, llama.cpp CUDA or PyTorch CUDA install paths.
  • Added project-tools mode switcher: Установка / Проверка.

Changed

  • Split the previous mixed OCR Brick window into separate local and paid OCR root windows.
  • Reworked local/paid OCR dialogs into logical visual blocks: engine/model/access, preprocessing, local 2-pass, prompt/parameters, postprocessing, output and advanced.
  • Replaced the visible preprocessing control field with a single engine-aware profile selector; detailed knobs are hidden until Ручной is selected.
  • Reworked model selectors into wide dropdowns with pin/unpin/update controls and curated defaults.
  • Shortened noisy operation labels and tightened the main root-menu row spacing.
  • Moved coordinate-table extraction to root as Таблицы.
  • Moved Tesseract install into Инструменты проекта -> Установка and marked it as recommended.
  • Moved project checks, model refreshes and OCR smoke tests into Инструменты проекта -> Проверка.
  • Updated install/build scripts for the current portable runtime layout and optional OCR engines.
  • Updated DEV Markdown PDF output to support project docs/PDF, beside-source output and external target trees.
  • Updated vision OCR prompt handling to reduce decorative Markdown and preserve source text more strictly.
  • Updated Gemini OCR calls to pass service_tier=flex when Flex is selected and use longer timeouts for Flex queueing.
  • Replaced stale user guides with current root-window documentation and moved implementation history into this changelog.
  • Added docs/GEMINI_BATCH_FLEX.md with Gemini Standard/Flex/Batch usage, GUI mapping, artifact paths and the verified Вопрос 239.pdf dry-run.
  • Updated GitHub README files to describe OCR Brick root windows, current local/API OCR engines, Gemini Batch/Flex and one-page quality/requisites workflow instead of the legacy Workbench-first flow.
  • Added install documentation for optional OCR acceleration entries [11]-[17], including practical RTX 50xx/40xx/Vulkan/CPU choices and the distinction between Real-ESRGAN, Surya, llama.cpp and heavy PyTorch CUDA benchmark payloads.

Removed

  • Removed Paddle OCR from UI, install paths and project cleanup expectations.
  • Hid/removed legacy Python OCR Workbench from the visible root navigation.
  • Removed stale config/keys initialization from folder setup.
  • Removed legacy mixed root OCR actions in favor of the new local/paid OCR windows.

Verified

  • NiceGUI smoke test passes after UI changes.
  • Python compile checks pass for updated UI/service files.
  • Yandex OCR smoke passed with configured key/folder.
  • xAI model-list and OCR smoke passed.
  • Gemini Batch dry-run prepared 3 JSONL requests for Вопрос 239.pdf pages 6-8 without remote submission.
  • Tesseract OCR Brick runs were smoke-tested on 10_Самарская_Тольятти.pdf and 1_Псковская_Псковский.pdf.
  • Numeric check matched key Pskov identifiers with local locator plus xAI verification.

Cleanup

  • Legacy implementation plans, handoff briefs and generated benchmark/log/cache artifacts were removed from the working tree after this changelog captured the relevant decisions.
  • Source test package and smoke-run workspaces were removed from the portable project cleanup set.