PDFusion: Local PDF Analyzer & Markdown Converter 🚀

PDFusion is a comprehensive PDF analysis application that extracts text, images, and tables from uploaded PDFs and outputs everything in clean, structured Markdown format. It uses a local Python backend with FastAPI and utilizes local AI models via LiteLLM for advanced OCR.

🏔️ Technical Architecture

Core Stack

• Frontend: React 18+ with TypeScript (Vite)
• Backend: FastAPI (Python 3.10+)
• Database: Postgres (local or homelab) with SQLite auto-fallback
• AI/LLM: LiteLLM Proxy + Ollama (Local Vision Models)
• PDF Extraction: PyMuPDF (fitz) + Pandas (Tables)

System Architecture

┌─────────────────┐    ┌──────────────────┐    ┌─────────────────┐
│   React App     │───▶│   FastAPI        │───▶│   LiteLLM Proxy │
│   (Frontend)    │    │   (Backend)      │    │   (Ollama)      │
└─────────────────┘    └──────────────────┘    └─────────────────┘
                                │                          │
                                ▼                          ▼
                         ┌──────────────┐         ┌──────────────┐
                         │  PostgreSQL/ │         │  Local       │
                         │  SQLite      │         │  AI Models   │
                         └──────────────┘         └──────────────┘

---

🚀 Getting Started

Prerequisites

• Node.js (v18 or later)
• Python (v3.10 or later)
• Ollama (Running locally with gemma4 models)
• Postgres (Optional, will fallback to local SQLite)

1. Backend Setup

Navigate to the root directory and set up the Python environment:

# Create a virtual environment
python -m venv venv

# Activate the virtual environment
# On Windows:
.\venv\Scripts\activate
# On macOS/Linux:
source venv/bin/activate

# Install dependencies
pip install -r server/requirements.txt

2. Frontend Setup

Install the Node.js dependencies:

npm install

3. AI Model Setup (Ollama / LiteLLM)

Ensure Ollama is installed and your vision model is pulled. If using a LiteLLM Proxy (recommended for homelab setups), ensure the endpoint is accessible.

ollama run gemma4

4. Configuration

Create a .env file in the root directory (based on .env.example).

Tip

PDFusion is configured to bypass self-signed SSL certificates for litellm.local homelab proxies automatically.

# API URL for the local backend
VITE_API_BASE_URL=http://localhost:8000

# PostgreSQL Database URL
DATABASE_URL=postgresql://user:password@localhost:5432/pdfusion

# LiteLLM Proxy / Ollama Configuration
VISION_MODEL_ID=ollama/gemma4
LITELLM_API_BASE=https://litellm.local
LITELLM_API_KEY=sk-local-proxy

---

🏃 Running the Application

Start the Backend Server

npm run server

This starts the FastAPI server at http://localhost:8000. It will auto-mount uploads/ and outputs/ for image serving.

Start the Frontend

npm run dev

The application will be accessible at http://localhost:5173.

---

🎨 'Pro' Viewer Features

The results dashboard now includes a professional-grade Markdown viewer inspired by advanced editors:

• GFM Support: Full rendering of complex data tables and strikethroughs.
• Math/KaTeX: Support for mathematical formulas extracted from documents.
• Image Intelligence: Automatic URL rewriting ensures embedded images are served correctly from the local backend.
• Local OCR: Full text extraction displayed clearly beneath every identified image.

---

📁 Project Structure

├── server/                # Python FastAPI Backend
│   ├── src/               # Processing logic (PDF, Vision, Table)
│   ├── uploads/           # Local storage for raw PDFs
│   ├── outputs/           # Local storage for extracted images/assets
│   ├── main.py            # FastAPI entry point
│   ├── models.py          # Database models (SQLAlchemy)
│   └── requirements.txt   # Python dependencies (Pandas, Pillow, etc.)
├── src/                   # React Frontend (Vite)
│   ├── components/        # UI Components & Icons
│   ├── services/          # API Client Layer
│   └── pages/             # Dashboard & Results Views
└── package.json           # Frontend dependencies (ReactMarkdown, RemarkGfm)

---

🎯 Key Features

• Zero Cloud Leakage: No data leaves your machine or your local network.
• Intelligent Table Reconstruction: Uses Pandas for high-fidelity table parsing.
• Private Vision OCR: High-performance image-to-text via local LiteLLM proxy.
• Export Ready: Clean, standardized Markdown output.

---

⚖️ License

This project is licensed under the MIT License.