PrettyPipeline

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

PDF  →  embedded text (digital PDFs) or Unlimited-OCR (scans)
                              ↓
        GPT-5.4 nano (OpenAI)  →  JSON + needs_review + _meta

Why this is cheap

Cloud OCR APIs charge per page. PrettyPipeline does vision work locally with baidu/Unlimited-OCR (MIT, 3B params, 32K context) when a page has no usable embedded text.

The only paid call is GPT-5.4 nano on document text (~$0.20 / 1M input tokens), prompted with your JSON schema.

Digital PDFs use embedded text automatically (fast, accurate). OCR runs per-page only for scans, or for everything with --force-ocr.

Nulls, uncertain fields, values not found in source text, and garbled OCR are flagged in needs_review (one reason per field).

Install

Python 3.10–3.13 (3.12 recommended).

Digital PDFs only (no local OCR, smallest install):

pip install prettypipeline-ocr

Full install (local OCR for scans and --force-ocr):

pip install prettypipeline-ocr[ocr]

The CLI is prettypipeline. From this repo:

python3.12 -m venv .venv && source .venv/bin/activate
pip install -e ".[ocr,dev]"

NVIDIA Linux — install a CUDA PyTorch wheel first if needed:

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

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

Docker (CUDA):

docker build -f Dockerfile.cuda -t prettypipeline:cuda .
docker run --gpus all -e OPENAI_API_KEY=sk-... \
  -v "$PWD/data:/data" prettypipeline:cuda \
  run /data/invoice.pdf --schema /data/schema.json -o /data/out.json

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

Batch:

prettypipeline batch ./invoices/*.pdf \
  --schema examples/invoice.schema.json \
  --output-dir ./out/ \
  --summary ./out/summary.json

Schema examples

--schema is any JSON Schema — swap the file to extract a different document type.

CLI

prettypipeline run FILE.pdf --schema SCHEMA.json
  -o, --output PATH           write JSON (also printed)
  --ocr-only                  skip OpenAI step
  --force-ocr                 always OCR (skip embedded text)
  --include-ocr-text          include raw text in output
  --device cuda|mps|cpu       override auto-detect
  --dpi N                     PDF raster DPI (default 300)
  --max-length N              OCR cap (8192 MPS, 32768 CUDA/CPU)
  --model NAME                LLM model (default gpt-5.4-nano)
  --base-url URL              OpenAI-compatible API endpoint

prettypipeline batch FILES... --schema SCHEMA.json --output-dir DIR
  --summary PATH              optional batch summary JSON

Environment variables:

Output

{
  "data": {
    "date": "January 25, 2016",
    "vendor": "DEMO - Sliced Invoices",
    "invoice_number": "INV-3337",
    "total": 93.5
  },
  "needs_review": [],
  "source": "pdf_text",
  "_meta": {
    "version": "0.3.0",
    "elapsed_ms": 4200,
    "pages": 1,
    "device": "mps",
    "model": "gpt-5.4-nano",
    "token_usage": {
      "prompt_tokens": 1200,
      "completion_tokens": 180,
      "total_tokens": 1380
    }
  }
}

Programmatic use

from prettypipeline import run

result = run(
    "invoice.pdf",
    "examples/invoice.schema.json",
)
print(result.data)
print(result.needs_review)
print(result.meta)

Lower-level imports: extract_text, structure, needs_review.

Hardware

On MPS, default --max-length is 8192 (input + generation). CUDA/CPU default is 32768.

Development

pip install -e ".[ocr,dev]"
pytest
python -m build && twine check dist/*

License

MIT. Unlimited-OCR is also MIT.

Sample invoice: Sliced Invoices.