Sift — Digital Document Sifter

  ╔══════════════════════════════════════╗
  ║    ███████╗██╗███████╗████████╗      ║
  ║    ██╔════╝██║██╔════╝╚══██╔══╝      ║
  ║    ███████╗██║█████╗     ██║         ║
  ║    ╚════██║██║██╔══╝     ██║         ║
  ║    ███████║██║██║        ██║         ║
  ║    ╚══════╝╚═╝╚═╝        ╚═╝         ║
  ║       Digital Document Sifter        ║
  ╚══════════════════════════════════════╝

Sift finds the images and videos in a folder that actually contain text, and pulls them out so you don't have to scroll past 10,000 photos to find them.

What it's for

• Old phone backups. Years of WhatsApp/Telegram media dumped into one folder — screenshots of addresses, OTPs, boarding passes, and ID scans buried among memes and photos. Point Sift at the folder and it surfaces the ones with text.
• Downloads folder archaeology. Random screenshots you saved "for later" — whiteboard photos, error messages, code snippets, receipts — mixed in with everything else. Sift separates them out.
• Screen recordings and video clips. Tutorials, lecture recordings, bug-report clips — find the ones with on-screen text without re-watching every file.
• General declutter. Anywhere photos and "documents disguised as photos" have piled up together.

Run it, walk away, come back to a SIFT_docs/ subfolder with the text-bearing files sorted out. Everything else stays exactly where it was.

How it works (briefly)

Sift OCRs every image (and, in video mode, sampled frames of every video) with Tesseract. Anything with detectable text gets moved into SIFT_docs/ in the same folder it was found in. It auto-scales to use ~70% of your CPU cores so it doesn't lock up your machine while it runs. Video mode requires ffmpeg on PATH.

Limitations

OCR isn't perfect — read this before relying on it for anything important:

• Misses low-contrast, stylized, or handwritten text
• Sensitive to image rotation and resolution
• English only by default (language-pack dependent)
• No image preprocessing, so noisy photos or photos-of-screens can produce false negatives
• Video mode only samples frames every 1–5 seconds — brief on-screen text can be missed entirely

Sift gets you most of the way there. It doesn't replace manually skimming the untouched files if completeness actually matters.

Install

Binary download (recommended — no dependencies)

Download the latest binary for your platform from the releases page:

# Linux/macOS
curl -fsSL https://github.com/laghab/sift/releases/latest/download/sift_linux_amd64.tar.gz | tar xz
./sift ~/photos

# Windows
# Download sift_windows_amd64.tar.gz from releases, extract, run sift.exe

Or via install script:

curl -fsSL https://github.com/laghab/sift/releases/latest/download/install.sh | bash

Go toolchain

go install github.com/laghab/sift/cmd/sift@latest

Legacy Python version (GUI only)

The Go rewrite is CLI-only. For the Tkinter GUI, use the Python version:

git clone https://github.com/laghab/sift.git
cd sift
pip install -e .
sift-gui

Usage

sift ~/photos                      # scan images (default: move text files)
sift ~/photos --copy               # copy instead of move
sift ~/photos --dry-run            # preview what would be moved
sift ~/photos --search invoice     # only match files containing "invoice"
sift ~/photos --json               # machine-readable JSON output
sift ~/photos --video              # scan videos instead (needs ffmpeg on PATH)
sift ~/photos --threshold 10       # require 10+ meaningful words
sift ~/photos --workers 4          # override CPU count

All flags can be combined. For example:

sift ~/WhatsAppBackup --copy --search "address" --json

License

MIT — see LICENSE.