PrettyPipeline

Turn PDFs into structured JSON. OCR runs on your GPU. Field extraction uses a cheap cloud LLM. You define the schema.

PDF  →  PyMuPDF (pages → images)  →  Unlimited-OCR (local)  →  raw text
                                                              ↓
                                    GPT-5.4 nano (OpenAI)  →  JSON + needs_review

Why this is cheap

Cloud OCR APIs charge per page. PrettyPipeline does the expensive vision work locally with baidu/Unlimited-OCR (MIT, 3B params, 32K context). A whole PDF is parsed in one infer_multi() pass — no page-chunking.

The only paid call is GPT-5.4 nano on that text (~$0.20 / 1M input tokens), prompted with your JSON schema. You pay for a small text completion, not for pixels.

Nulls, model-marked uncertain fields, values not found in the source text, and OCR that looks garbled are flagged in needs_review instead of being silently accepted.

Digital PDFs — embedded text is used when available (fast, accurate). OCR runs only for scans or with --force-ocr.

Install

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

pip install prettypipeline-ocr

The CLI name is prettypipeline. From this repo:

python3.12 -m venv .venv
source .venv/bin/activate
pip install -e .

NVIDIA Linux — install a CUDA wheel of PyTorch first if pip gave you CPU-only:

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

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

Quick start

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

Demo invoice in this repo:

prettypipeline run tests/fixtures/sample_invoice.pdf \
  --schema examples/invoice.schema.json -o out.json

Digital PDFs use embedded text automatically. Use --force-ocr to always run Unlimited-OCR.

Schema

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

{
  "type": "object",
  "properties": {
    "date": { "type": ["string", "null"] },
    "vendor": { "type": ["string", "null"] },
    "line_items": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "description": { "type": ["string", "null"] },
          "amount": { "type": ["number", "null"] }
        }
      }
    },
    "total": { "type": ["number", "null"] }
  }
}

See examples/invoice.schema.json for a fuller invoice example.

CLI

prettypipeline run FILE.pdf --schema SCHEMA.json
  -o, --output PATH     write JSON to a file (also printed)
  --ocr-only            skip the OpenAI step
  --force-ocr           always OCR, skip embedded PDF text
  --device cuda|mps|cpu override auto-detect (cuda → mps → cpu)
  --dpi N               PDF raster DPI (default 300)
  --max-length N        OCR cap (default 2048 on MPS, 32768 on CUDA/CPU)

OPENAI_API_KEY is required unless you pass --ocr-only. It is never hardcoded.

Output

{
  "data": {
    "date": null,
    "vendor": "DEMO - Sliced Invoices",
    "bill_to": "Test Business",
    "line_items": [],
    "total": null
  },
  "needs_review": [
    { "field": "date", "reason": "null" },
    { "field": "total", "reason": "null" }
  ],
  "ocr_text": "…"
}

Hardware

License

MIT. Unlimited-OCR is also MIT.

Sample invoice fixture: Sliced Invoices.