Docsort

Fully offline, privacy-first document sorting daemon. Watches an inbox folder for scanner-dropped documents (PDF, TIFF, images), runs OCR, extracts metadata via a local LLM, and files them with deterministic names.

Features

• Offline & private — all processing stays on your machine (Ollama for LLM, ocrmypdf for OCR)
• German + English OCR via ocrmypdf with sidecar text extraction
• LLM metadata extraction via local Ollama (document type, date, sender, subject, etc.)
• Deterministic filenames built from extracted metadata with collision handling
• SHA256 deduplication — duplicates are detected and moved aside
• Crash-safe pipeline — SQLite-backed state machine with recovery on restart
• Web dashboard — local web UI at localhost:8484 with pipeline status, document list, and one-click retry/re-OCR/re-extract
• CLI management — status, retry, re-OCR, re-extract, and garbage collection commands

Prerequisites

• ocrmypdf with Tesseract language packs (deu, eng)
• Ollama running locally with a model pulled (default: llama3.2)
• Rust toolchain (for building)

Build

cargo build --release

Usage

Start the daemon

docsort daemon

This watches docsort/inbox/ for new files and processes them through the pipeline:

inbox/ → stability check → normalize to PDF → OCR → text cleanup → LLM extract → deterministic rename → output/

CLI options

docsort --base-dir ./my-docs daemon    # custom base directory
docsort --ollama-url http://host:11434 daemon
docsort --llm-model llama3.2 daemon
docsort --ocr-languages deu+eng daemon

Web dashboard

The daemon starts a web dashboard at http://localhost:8484 with:

• Pipeline overview — card widgets showing total, processed, failed, and in-progress counts
• Document list — filterable by stage (queued, processed, failed, etc.)
• Document detail — all extracted metadata (sender, type, date, title, etc.) with action buttons
• One-click actions — retry, re-OCR, and re-extract directly from the UI
• Auto-refreshes every 5 seconds

Management commands

docsort status                  # show pipeline counts and recent errors
docsort retry --id 42           # retry a failed document
docsort retry --all             # retry all failed documents
docsort reocr --id 42           # re-run OCR on a document
docsort reextract --id 42       # re-run LLM extraction (keeps OCR text)
docsort gc --max-age-days 7     # clean up old tmp/ directories

Folder structure

docsort/
  inbox/          # drop scanned files here
  processing/     # in-flight documents
  tmp/            # per-document work artifacts
  output/         # final OCR'd PDFs with deterministic names
    _duplicates/  # SHA256 duplicates
  failed/         # failed documents with diagnostic bundles
  db/
    docsort.sqlite

License

MIT