PDF Classification & Extraction Platform (Flask)

A clean Flask starter architecture for a PDF intelligence pipeline based on your flowchart:

• PDF upload endpoint
• Pre-processing (PyMuPDF + EasyOCR fallback placeholder)
• Main bucket classifier (loads externally trained checkpoints; heuristics fallback)
• Sub-bucket classifier
• Key info extractor (spaCy NER + regex placeholder)
• Storage/grouping engine
• Polished web UI

Project Structure

my_second_python/
├── app/
│   ├── __init__.py
│   ├── config.py
│   ├── models/
│   │   └── document.py
│   ├── routes/
│   │   └── main.py
│   ├── services/
│   │   ├── classify.py
│   │   ├── extract.py
│   │   ├── grouping.py
│   │   ├── mlops.py
│   │   └── preprocess.py
│   ├── static/
│   │   ├── css/style.css
│   │   └── js/app.js
│   └── templates/
│       ├── base.html
│       ├── index.html
│       └── result.html
├── data/
│   ├── processed/
│   └── uploads/
├── requirements.txt
└── run.py

Run Locally

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python run.py

Then open http://127.0.0.1:5000

External training (e.g. Kaggle)

Fine-tune distilbert-base-multilingual-cased on data/training/train.jsonl / validation.jsonl, then place checkpoints under models/main_bucket_classifier/ and models/sub_bucket_classifier/. See docs/kaggle-training.md for labels, layout, and wiring.

Next Up (Production)

• Plug real models into app/services/classify.py (ONNXRuntime INT8)
• Add PyMuPDF + EasyOCR logic in app/services/preprocess.py
• Replace in-memory grouping with SQLite persistence
• Add feedback endpoint to send correction events to retraining queue