title	PDFSystem: PB-Scale PDF Processing Pipeline
emoji	🚀
colorFrom	green
colorTo	purple
sdk	gradio
sdk_version	6.12.0
app_file	app.py
pinned	false
license	apache-2.0
short_description	PDF to Markdown pipeline with ML-powered routing

PDFSystem for MNBVC

PB-scale PDF → Pretraining Data Pipeline
FinePDFs-inspired architecture for Chinese-heavy, mixed-quality PDFs

---

🚀 Quick Links

Platform	Link	Description
Live Demo	🤗 HF Spaces	Upload PDF and try the pipeline instantly
Source Code	GitHub	Full source code and documentation

---

✨ Features

• 🧠 ML-Powered Routing: XGBoost classifier (124 features) routes PDFs to optimal backend
• ⚡ Fast Path: PyMuPDF extraction for text-ok documents (~10ms/page)
• 📊 Quality Scoring: ModernBERT-large OCR quality assessment [0-3 scale]
• 🔍 Visual Debug: Page preview with extracted bbox overlays
• 📦 Modular Design: Stateless, backend-agnostic pipeline components

---

🎯 Current Status

Component	Status	Description
Stage-A Router	✅ Ready	XGBoost binary classifier with 124 PyMuPDF features
MuPDF Parser	✅ Ready	Fast extraction for clean-text PDFs
OCR Quality Scorer	✅ Ready	ModernBERT-large regression model
Stage-B Router	✅ Ready	LayoutDocument → PIPELINE / VLM / DEFERRED
Layout Analyser	✅ Ready	DocLayout-YOLO + PP-DocLayoutV3 (dual backend)
Pipeline Parser	✅ Ready	Region-level OCR via RapidOCR
VLM Parser	✅ Ready	MinerU 2.5 Pro (magic-pdf) for complex pages
Unified CLI	✅ Ready	pdfsys run -c config.yaml --stages ...
Annotation UI	✅ Ready	pdfsys annotate — PDF labeling + layout overlay

---

🏃 Quick Start

Option 1: Online Demo (Fastest)

Visit Hugging Face Spaces and upload a PDF — no installation required.

Option 2: Local Development

# 1. Install uv package manager
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Clone and setup
git clone https://github.com/MIracleyin/pdfsystem_mnbvc.git
cd pdfsystem_mnbvc
uv sync

# 3. Download router weights (257 KB, one-time)
python -m pdfsys_router.download_weights

# 4. Run interactive demo
python app.py
# Open http://localhost:7860

Option 3: Batch Processing

python -m pdfsys_bench \
  --pdf-dir /path/to/pdfs \
  --out results.jsonl \
  --markdown-dir ./extracted

---

🏗️ Architecture

                    ┌─────────────────┐
   PDF Input  ───►  │  Stage-A Router │  XGBoost (124 features)
                    │  (Implemented)  │  ~10ms per PDF
                    └────────┬────────┘
                             │ ocr_prob
           ┌─────────────────┼─────────────────┐
           ▼                 ▼                 ▼
      ┌─────────┐      ┌──────────┐      ┌─────────┐
      │  MUPDF  │      │ PIPELINE │      │   VLM   │
      │  (Fast) │      │  (OCR)   │      │(Complex)│
      └────┬────┘      └──────────┘      └─────────┘
           │
           ▼
   ┌─────────────────────────────────────┐
   │  ExtractedDoc: Markdown + Segments  │
   └─────────────────────────────────────┘
           │
           ▼
   ┌─────────────────────────────────────┐
   │  Quality Scorer (ModernBERT-large)  │
   │  Score: [0, 3]                      │
   └─────────────────────────────────────┘

---

📦 Workspace Packages

Package	Purpose	Dependencies
pdfsys-core	Shared types, schemas, layout cache	stdlib only
pdfsys-router	Stage-A/Stage-B routing decisions	pymupdf, xgboost, pandas, sklearn
pdfsys-parser-mupdf	Fast PyMuPDF extraction	pymupdf
pdfsys-bench	Evaluation harness + quality scorer	torch, transformers
pdfsys-layout-analyser	DocLayout-YOLO / PP-DocLayoutV3 detection	doclayout-yolo, transformers
pdfsys-parser-pipeline	Region-level OCR via RapidOCR	rapidocr-onnxruntime
pdfsys-parser-vlm	MinerU 2.5 Pro VLM extraction	magic-pdf
pdfsys-cli	Unified CLI + YAML config + annotation UI	pyyaml

---

📊 Benchmark Results

OmniDocBench-100 Dataset:

Backend Split:    mupdf=70    pipeline=30
Avg OCR Prob:     mupdf=0.034  pipeline=0.634
Extraction:       70 success   0 errors
Quality Score:    avg=1.71     min=0.39   max=2.73
Timing:           router=49ms  extract=7ms  quality=3.6s

---

🎨 Demo Interface

The Gradio demo provides:

• 📤 PDF Upload: Drag-and-drop or click to upload
• 📈 Routing Info: OCR probability, selected backend, page count
• 🖼️ Page Preview: First page with colored bbox overlays
• 📝 Markdown Output: Extracted text content
• 📋 Segment Table: Block-level extraction details
• 🔧 Feature View: Selected router features
• 📄 Raw JSON: Complete pipeline output
• ⭐ Quality Score: Optional ModernBERT scoring

---

📚 Documentation

Document	Description
docs/PRD.md	Product Requirements & Architecture Rationale
docs/ROADMAP.md	Implementation Timeline & Milestones
CONTRIBUTING.md	Development Guidelines & Commit Conventions
demo/README.md	Demo-specific Documentation

---

💻 Development

Data Structures

Router Output:

@dataclass
class RouterDecision:
    backend: Backend          # MUPDF | PIPELINE | VLM | DEFERRED
    ocr_prob: float           # P(needs OCR) [0, 1]
    num_pages: int
    is_form: bool
    features: dict            # 124-dim feature vector

Parser Output:

@dataclass(frozen=True)
class ExtractedDoc:
    sha256: str
    backend: Backend
    segments: tuple[Segment, ...]
    markdown: str
    stats: dict

CLI Reference

# Download router weights
python -m pdfsys_router.download_weights

# Run benchmark
python -m pdfsys_bench \
  --pdf-dir PATH \
  --out results.jsonl \
  --no-quality          # Skip quality scoring

---

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

---

📄 License

This project is licensed under the Apache License 2.0.

---

Built with ❤️ for the MNBVC corpus project