Team Name: Fruits of Binary Tree

Challenge: Adobe Hackathon Round 1A

This project delivers automatic extraction of semantic outlines from PDF files—titles and document headings, with their hierarchy and page numbers. It is optimized for offline, CPU-only environments, and is fully containerized (Docker, AMD64) for seamless deployment and reproducibility in real-world or hackathon scenarios.

Key Features

• Smart Section Heading Detection: Extracts headings using a hybrid of font size, position, layout clustering, and transformer-based semantic classification (DistilBERT-ONNX).
• Hierarchical JSON Output: Provides strong structure: title, H1/H2/H3... headings with associated page numbers.
• CPU-Optimized and Under 200MB: Quantized ONNX transformer model ensures rapid, memory-light inference.
• Batch, Hands-free Processing: Scans all PDFs from /app/input/, writes [filename].json for each into /app/output/.
• Runs Completely Offline: Zero internet required after build—perfect for secure/censored or constrained environments.
• DevOps Ready: Packaged for Docker, supporting host folder mounts and AMD64 CPU image.

Quickstart: Build & Run

1. Place your input PDFs

• Put your .pdf files in a host folder called input at the same directory as your Dockerfile.

2. Build the Docker Image

docker build --platform linux/amd64 -t pdf-outline-extractor:latest .

3. Run the Container

docker run --rm \
    -v $(pwd)/input:/app/input \
    -v $(pwd)/output:/app/output \
    --network none \
    pdf-outline-extractor:latest

• All JSON output will be in ./output/, one file per input PDF.

Project Workflow Overview

Core Tech & Model Choices

• PyMuPDF (fitz): Fast PDF layout/geometry parsing—including font/style/position info.
• ONNX Runtime: Executes a quantized DistilBERT for heading/paragraph discrimination.
• Python 3.10 (slim base): Fast, minimal footprint.
• No external API/model downloads—everything included during build.