Adobe Round 1B – Intelligent Document Section Extractor

📝 Overview

This project addresses Challenge 1B of the Adobe Document Intelligence task. The goal is to automatically extract the most relevant sections and sub-sections from a set of input documents (PDFs) based on a provided persona and job-to-be-done (JTBD) description. The output is a structured JSON containing metadata, ranked sections, and refined sub-section analysis.

---

📂 Directory Structure

Adobe_Round_1b/
├── input/                     # Folder for input PDFs and persona
│   ├── *.pdf
│   └── challenge1b_input.json
├── output/
│   └── challenge1b_output.json
├── models/                    # Pre-downloaded FLAN-T5/MT5 model files
├── main.py                    # Entry-point script for processing
├── src/                     # Helper modules
│   └── input_loader.py
│   └── pdf_extractor.py
│   └──relevance_ranker.py
│   └──text_refiner_llm.py
├── Dockerfile
├── requirements.txt
├── approach_explanation.md
└── README.md

---

⚙️ Execution Instructions (Docker)

Here’s your Docker usage instructions formatted and highlighted in Markdown:

⸻

🐳 Docker Instructions

🔧 1. Build the Docker Image

docker build -t adobe_round_1b .

This will build the Docker image using the provided Dockerfile.

⸻

🚀 2. Run the Docker Container

Assuming your input PDFs and JSON are in the input/ folder, run:

Adobe_Round_1b/
├── input/                     # Folder for input PDFs and persona
│   ├── *.pdf
│   └── challenge1b_input.json

docker run --rm \
  -v $(pwd)/input:/app/input \
  -v $(pwd)/output:/app/output \
  adobe_round_1b

✅ This will mount your local input/ and output/ folders to the container and run the pipeline.

📄 After execution, the output will be saved as:

output/challenge1b_output.json

Here is a cleanly structured and well-formatted version of your content in Markdown:

⸻

📤 Output Format

The generated output JSON follows the structure outlined below:

1. 🧾 Metadata

{
  "input_documents": [...],
  "persona": "...",
  "job_to_be_done": "...",
  "processing_timestamp": "..."
}

2. 📚 Extracted Sections

[
  {
    "document": "doc1.pdf",
    "page_number": 3,
    "section_title": "Experience Overview",
    "importance_rank": 1
  },
  ...
]

3. 🧠 Sub-section Analysis

[
  {
    "document": "doc1.pdf",
    "refined_text": "Detailed sub-section relevant to job",
    "page_number": 3
  },
  ...
]

✅ The output strictly follows the required format defined in challenge1b_output.json.

⸻

📦 Deliverables • approach_explanation.md – Methodology and model design explanation (300–500 words). • Dockerfile – Container setup for offline, CPU-only execution. • README.md – This file with setup instructions, usage, and references. • output/challenge1b_output.json – Final output generated from sample input documents.

⸻

⚖️ Evaluation Criteria Mapping • Section Relevance (60 points) → Uses embedding similarity + persona/job scoring with intelligent ranking. • Sub-section Relevance (40 points) → Refined sub-section summaries using local mT5 model. • Execution Constraints • ✅ CPU-only • ✅ Model size ≤ 1GB • ✅ Offline inference using pre-downloaded model weights. • Output Format • ✅ Structured JSON output • ✅ Includes metadata, extracted sections, and sub-section summaries.

⸻

📄 Refer To • approach_explanation.md for: • Document extraction pipeline • Ranking mechanism • LLM-based sub-section refinement • Offline execution setup with local models

⸻

👨‍💻 Authors

👥 TeamIronMan – Adobe Hackathon 2025

This solution was collaboratively built as part of the Adobe Hackathon 2025 by TeamIronMan 🚀:

Jitendra Kolli : https://github.com/jitendra-789

Prasanth Kumar : https://github.com/prasanth1221