🎯 Adobe India Hackathon 2025 – Connecting the Dots

"Rethink Reading. Rediscover Knowledge."

Welcome to our submission for the Adobe India Hackathon 2025.
Our project transforms static PDFs into dynamic, intelligent documents — capable of understanding structure, surfacing insights, and connecting dots across knowledge sources — all offline.

🧠 Hackathon Overview

The hackathon consists of two technical rounds:

Round	Focus
🟦 Round 1A	Extract structured document outlines (Title, H1–H3)
🟩 Round 1B	Surface sections relevant to a specific persona

✨ Features Summary

Feature	Status
Extract Title, H1–H3 Headings	✅ Implemented
Font-size & Layout-based Logic	✅ Implemented
Structured JSON Output	✅ Yes
Fully Offline (No Web Access)	✅ Yes
CPU-Only Execution	✅ Yes
Docker Support	✅ Yes
Sample Input/Output Provided	✅ Yes
Multilingual PDF Support	⚙️ Planned

🟦 Round 1A – PDF Outline Extractor

📌 Objective

Automatically extract the Title, and H1 / H2 / H3 headings with their corresponding page numbers from any PDF (≤ 50 pages), and output a valid JSON as per Adobe’s spec.

✅ Sample Output Format

{
  "title": "Understanding AI",
  "outline": [
    { "level": "H1", "text": "Introduction", "page": 1 },
    { "level": "H2", "text": "What is AI?", "page": 2 },
    { "level": "H3", "text": "History of AI", "page": 3 }
  ]
}

📂 Folder Structure

Round1A/
├── app/
│   ├── input/                   # Input PDFs placed here (Docker-mounted)
│   └── output/                  # Output JSONs written here (Docker-mounted)
├── src/
│   ├── main.py                  # Entrypoint for processing all PDFs
│   ├── extractor.py             # Extracts title, H1, H2, H3 using font sizes
│   ├── utils.py                 # Helpers for reading PDF, writing JSON
│   └── config.py                # Thresholds/configs for heading detection
├── requirements.txt             # pdfplumber + PyMuPDF
├── Dockerfile                   # CPU-only, offline, AMD64-compliant
├── generate_dummy_pdf.py        # (Optional) Generate test PDFs with headings
├── sample.pdf                   # (Optional) A test input PDF
└── sample.json                  # (Optional) Expected output for sample.pdf

🧰 Tech Stack

Tool/Library	Use Case
pdfplumber, PyMuPDF	Parsing PDF text + layout
sentence-transformers	Semantic relevance ranking (R1B)
scikit-learn, numpy	Similarity scoring, vector ops
transformers (optional)	Summarization (R1B, optional)
Docker	CPU-only, offline deployment
Python 3.10+	Primary language

📌 All tools meet the offline, lightweight, and CPU-compliant constraints.

⚙️ How It Works (Round 1A)

1. Place .pdf files in the app/input/ folder.

2. Run the program:

python src/main.py

3. It will:

• ✅ Extract title and headings (H1–H3)
• ✅ Save output JSON in app/output/ with the same filename

📦 Dependencies

pdfplumber==0.10.2  
PyMuPDF==1.23.1

Install locally with:

pip install -r requirements.txt

🧠 Refined Heuristics for Heading Detection

Our system simulates how a human visually parses a document — not just scans font sizes. Here’s how we do it:

#️⃣	Heuristic Rule	Signal Type
1	Title must appear in top 15–25% of page 0, based on Y-position	Layout + Visual
2	Font sizes are ranked dynamically per document (largest = H1)	Font Heuristic
3	Headings must be ≤ 3 lines and ≤ 120 characters	Content Filter
4	ALL CAPS or bold text ≠ heading unless layout supports it	Visual + Context
5	Boost semantic phrases like "Goals", "Summary", "Appendix"	NLP + Semantics
6	Skip content inside tables, forms, QA blocks	Layout Heuristic
7	Preserve section numbers like 2.1 Mission — never split/truncate	Content Rule
8	Merge multi-line headings only if alignment and spacing match	Visual Merge
9	Prefer blocks with white space padding above and below	Structural Cue
10	Skip paragraph-like blocks that appear bold but aren’t section-defining	Noise Filter
11	Repeated patterns (e.g., "Step 1", "Phase X") hint heading structure	Pattern Learning
12	Promote early headings when no clear H1 exists, to prevent outline starvation	Recovery Logic
13	Preserve all symbols/punctuation: no normalization (Goals:, not Goals)	Output Policy
14	Indented headings are allowed if visually distinct & top-aligned	Layout Analysis
15	Output must read like a Table of Contents, not just text with sizes	UX-Oriented Rule

🔍 View rules 16–35

#️⃣	Heuristic Rule	Signal Type
16	Ignore headers/footers repeated across pages	Layout Filter
17	Remove text with high frequency + small font across pages	Noise Control
18	Penalize left/right page margin-aligned content	Layout Heuristic
19	Titles with no sibling block nearby are considered isolated → boost score	Position Scoring
20	Headings often follow white space	White Space Rule
21	Pages with no detected headings: fallback to top font chunks	Recovery Logic
22	Prefer phrases with verbs/nouns over adjectives	NLP Patterning
23	Visually centered blocks on page 1 → strong title candidates	Title Heuristic
24	Avoid text with large line-height	Visual Check
25	Penalize headings with multiple font styles in one line	Mixed Font Check
26	Limit each heading level to ≤ 30% of total blocks	Balance Check
27	Stop at 50 pages even if file is larger	Constraint Rule
28	Avoid headings ending with ellipses/colons (unless list intro)	Punctuation Rule
29	Heading must be larger or bolder than adjacent text blocks	Contrast Rule
30	Emphasize blocks that appear only once across document	Rarity Boost
31	Prefer headings that appear top-to-bottom sequentially	Logical Flow
32	Allow H2s inside H1s if indentation + size are justified	Nested Rule
33	Penalize headings shorter than 3 characters	Min-Length Guard
34	Promote aligned blocks with white space above & followed by body text	Composite Cue
35	Use weighted ensemble of heuristics + layout scoring	Final Scoring

🟩 Round 1B – Persona-Driven Document Intelligence

🎯 Objective

Given:

• A user persona
• A task
• A set of PDFs

➡️ Surface and rank the most relevant sections, plus optionally summarize.

📥 Input/Output Example

{
  "metadata": {
    "persona": "Undergraduate Chemistry Student",
    "job": "Prepare for reaction kinetics exam",
    "documents": ["doc1.pdf", "doc2.pdf"],
    "timestamp": "2025-07-16T18:30:00Z"
  },
  "sections": [
    {
      "document": "doc1.pdf",
      "page": 4,
      "section_title": "Reaction Mechanisms",
      "importance_rank": 1
    }
  ],
  "subsections": [
    {
      "document": "doc1.pdf",
      "page": 4,
      "refined_text": "The SN1 reaction involves a two-step mechanism..."
    }
  ]
}

⚙️ Planned Modules (Round 1B)

File	Description
parser.py	Splits PDF into logical chunks
ranker.py	Ranks sections via semantic similarity
summarizer.py	Summarizes sections (optional)
main.py	Pipeline orchestrator

✅ Round 1B Constraints Coverage

Requirement	✅ Met
CPU-only	✅
Offline execution	✅
≤ 1GB model	✅
Runtime ≤ 60s	✅
Valid JSON Output	✅

🐳 Docker Setup (Round 1A)

docker build --platform linux/amd64 -t round1a-extractor .

docker run --rm \
  -v $(pwd)/app/input:/app/input \
  -v $(pwd)/app/output:/app/output \
  --network none round1a-extractor

👥 Team DCODERZ

Member	Role
Jahnavi	Lead Developer
Sahithi	Document Intelligence Engineer

📅 July 2025 🏁 Adobe India Hackathon – Connecting the Dots

📎 License

This project is licensed under the MIT License.

"We don’t just extract — we understand. We don’t just read — we connect." — Team DCODERZ