PrettyPipeline

Turn PDFs into structured JSON. OCR runs locally; field extraction uses a cheap cloud LLM.

Why this is cheap

Cloud OCR APIs charge per page. PrettyPipeline does the expensive vision work on your GPU with baidu/Unlimited-OCR (MIT, 3B params, 32K context — a whole PDF in one infer_multi() pass). The only API call is GPT-5.4 nano on the resulting text (~$0.20 / 1M input tokens), prompted with your JSON schema.

You pay for a small text completion, not for pixels.

PDF  →  PyMuPDF pages  →  Unlimited-OCR (local GPU)  →  raw text
                                                      ↓
                              GPT-5.4 nano (OpenAI)  →  JSON
                                                      ↓
                              validation             →  needs_review flags

Nulls, model-marked uncertain fields, and OCR that looks garbled are flagged for a human instead of being silently accepted.

Install

Python 3.10–3.13 (3.12 recommended). NVIDIA GPU (CUDA) or Apple Silicon (MPS). CPU works but is slow.

python3.12 -m venv .venv
source .venv/bin/activate
pip install prettypipeline-ocr
# or from this repo:
pip install -e .

On NVIDIA Linux, install a CUDA build of PyTorch first if the default wheel is CPU-only:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install -e .

The first run downloads baidu/Unlimited-OCR from Hugging Face (~6GB).

Usage

export OPENAI_API_KEY=sk-...
prettypipeline run invoice.pdf --schema examples/invoice.schema.json
prettypipeline run invoice.pdf --schema examples/invoice.schema.json -o out.json
prettypipeline run invoice.pdf --schema examples/invoice.schema.json --ocr-only
prettypipeline run invoice.pdf --schema examples/invoice.schema.json --device mps
prettypipeline run invoice.pdf --schema examples/invoice.schema.json --max-length 2048

--schema is any JSON Schema. Swap the file to extract a different document type — nothing is hardcoded.

Device is auto-detected (cuda → mps → cpu). Override with --device.

On Apple Silicon, Unlimited-OCR can start looping (it hardcodes CUDA). Use --max-length 2048 for short docs. NVIDIA CUDA is the model's native path.

Output

{
  "data": { "date": "…", "vendor": "…", "line_items": [], "total": 0 },
  "needs_review": [{ "field": "tax", "reason": "null" }],
  "ocr_text": "…"
}

License

MIT. Unlimited-OCR is also MIT.