VisiTexta

VisiTexta is a Windows desktop app that extracts text from images and PDFs and saves the result as Markdown.

It runs locally on your PC. No cloud OCR API is required.

This repo currently targets Windows behavior explicitly.

Who is this for?

• Students who want text from notes or scanned pages.
• Office users who need text from screenshots or PDFs.
• Anyone who wants simple OCR output in .md format.

What is new in 2.0.0

• Works reliably for both images and PDFs.
• Streams OCR text live in the app while processing.
• Auto-downloads the default model on first run if no model is installed.
• Uses a local runtime bundle in the release package.
• Adds an optional runtime profile selector for CPU-compatible or accelerated local inference.
• Produces cleaner OCR-first Markdown output.
• Adds calmer preview tabs for Original, OCR, Notes / Extract, and Export, plus a compact status bar for model, runtime, storage mode, and progress.
• Adds source-linked study notes plus Markdown, plain-text, and searchable text-based PDF note export options.
• Adds Extract mode with worker/company presets for invoices, receipts, table-to-CSV, meeting photos / whiteboards, and contract key points, including Markdown plus structured JSON and CSV where it fits.

Supported files

• PNG
• JPG / JPEG
• PDF

What you get

• A Markdown file saved next to your original file as file.ocr.md.
• If that file name already exists, VisiTexta saves file (ocr 2).md, file (ocr 3).md, and so on instead of overwriting anything.
• Live preview in the app while OCR runs.
• Notes mode can include page references that jump back to the preview image while the job stays loaded in the app.
• Extract mode can produce a readable Markdown summary plus structured JSON, and CSV when the chosen preset exposes row data.
• Notes PDF export stays text-based so the exported notes remain searchable; page references are preserved as text rather than embedded page-image links.

Workflow modes

• Exact OCR keeps the OCR-focused Markdown output path.
• Notes turns OCR pages into study notes with page references such as Source: p. 3, plus Markdown, text, and searchable note PDF export.
• Extract uses business-oriented presets and includes an Uncertainty / Verification section for fields that may need manual review.

Quick start (for normal users)

1. Download release 2.0.0.
2. Choose one package style:
3. For portable use, unzip the app and run VisiTexta.exe.
4. For installer use, run the Windows installer and launch VisiTexta from the installed app.
5. Drop an image or PDF into the app.

Portable vs installer behavior

Portable mode

• Portable mode is intended for an unpacked copy of the app.
• VisiTexta stores its own app data beside the executable in portable-data\.
• That includes: portable-data\settings.json portable-data\history.json portable-data\models\ portable-data\temp\ portable-data\pasted-inputs\
• No OS config directory is used while portable mode is active.
• Portable mode is selected automatically for unpacked copies outside common Windows install folders.
• You can also force portable mode by putting portable-data\ or visitexta-portable.txt beside VisiTexta.exe before first launch.

Installer mode

• Installer mode is intended for the normal Windows-installed app.
• VisiTexta stores settings, history, models, temp files, and pasted inputs under: %LOCALAPPDATA%\VisiTexta\
• This keeps the install folder clean and matches normal Windows app expectations.

What users see in the app

• Settings now shows the exact storage mode and the exact paths for settings, history, models, and temp files.
• Settings also shows the active local runtime profile: CPU compatible is the safe default. Auto prefers a compatible accelerated runtime when one is bundled and the PC looks compatible. Accelerated if available tries the accelerated runtime first and falls back cleanly if it cannot start.
• Acceleration changes speed only. OCR semantics stay tied to the same model, prompt, and preprocessing path.
• OCR output files are still written next to the source file, not inside the app-data folder.

First run behavior (important)

• If no supported OCR model is found, VisiTexta will start downloading the recommended default profile automatically.
• The default profile is GLM-OCR using GLM-OCR.Q4_K_M.gguf.
• This is normal and only happens on first setup (or if you removed supported models).
• Keep the app open until the download completes.
• Curated model downloads resume from existing partial .part files when Hugging Face supports ranged downloads.
• Curated model downloads are checksum-verified before they are accepted.

Why first output can feel slow

• The first word may take a while to appear.
• On the first page, the model is loading and preparing context.
• After that, output streams progressively.

In short: initial delay is expected, then text should start flowing.

Portable package layout

VisiTexta 2.0.0/
  VisiTexta.exe
  bin/
    accelerated/
      vulkan/
  resources/
  portable-data/

Model notes

• VisiTexta now uses an explicit curated model registry instead of treating arbitrary GGUF filenames as fully supported.
• GLM-OCR is the recommended default profile.
• Additional curated profiles include Qwen2-VL OCR 2B and Qwen2.5-VL 3B.
• Some curated models also need an mmproj file. If required, VisiTexta validates the download and fetches the companion mmproj automatically.
• Existing legacy model folders are still discovered during upgrades so older installs do not break abruptly.
• New downloads always go to the active primary storage location shown in Settings.
• Advanced settings still include an experimental custom download field for power users, but unlisted GGUF models are treated as best-effort only.
• For experimental custom downloads, enter a full owner/repo/file.gguf path. Repo-only auto-selection is reserved for the curated supported profiles.

Temp files and recovery

• Temporary OCR work files are kept in the app-managed temp folder and are cleaned on startup.
• If VisiTexta closes during a job, the interrupted job is kept in history and marked as failed on the next launch.
• Pasted images are stored in the active app-data location so retries and history stay predictable.

Troubleshooting

• Error about missing runtime CLI: Make sure bin/llama-mtmd-cli.exe and bin/llama-server.exe exist.
• Accelerated runtime is unavailable or falls back to CPU: Open Settings and switch back to CPU compatible, or leave the profile on Auto. Acceleration is optional and only affects speed.
• Error about missing model: Open Settings and download one of the curated profiles (or let the GLM-OCR auto-download finish).
• Error about missing mmproj: Re-run model download from Settings so companion files are fetched.
• Portable copy is using %LOCALAPPDATA% when you expected portable mode: Put portable-data\ or visitexta-portable.txt beside VisiTexta.exe, then launch it again.

For developers

From repo root:

cd app
npm install
npm run tauri:dev

Build release:

cd app
npm run build
npm run tauri:build

Release notes for packagers:

• Portable packages should include a sibling portable-data\ folder or visitexta-portable.txt marker so the mode is unambiguous even before first run.
• Installer packages should be installed normally; app data lives under %LOCALAPPDATA%\VisiTexta, not in the install directory.
• npm run tauri:build:installer builds the Windows installer bundles.
• npm run tauri:build:portable builds a no-bundle release executable, stages portable-data\, and creates a portable zip.
• npm run release:qa runs the release gate: frontend build, cargo check, warm benchmark gate, and cold benchmark gate.
• npm run benchmark:gate:warm and npm run benchmark:gate:cold compare benchmark runs against the checked-in baselines in app/benchmarks/baselines/.