Adobe India Hackathon 2025: Connecting the Dots

This repository contains the solution for Round 1 of the Adobe India Hackathon "Connecting the Dots".
The project is divided into two parts:

• Round 1a: A structural outline extractor for PDF documents.
• Round 1b: A persona-driven document intelligence system.

---

🧠 Approach

📄 Round 1a: Document Outline Extraction

The primary challenge in Round 1a is to extract a hierarchical outline (Title, H1, H2, H3) from a PDF document.
A simple approach based on font sizes is unreliable and discouraged.

✅ Our Approach:

A rule-based system that analyzes various text features to identify headings.

🔍 Text Extraction:

• We use PyMuPDF (fitz) for efficient extraction of text blocks and their properties.
• Chosen for its speed and precision, critical for the 10-second limit on a 50-page PDF.
• Extracted properties include:
  • Text content
  • Font size
  • Font name
  • Coordinates

⚙️ Feature Engineering:

For each text block, we analyze:

• Font Weight: Bold text suggests headings.
• Font Size: Larger fonts often indicate headings.
• Line Spacing: Headings usually have more spacing above/below.
• Text Length: Headings are typically shorter.
• Numbering: Patterns like 1.1, A., i. are strong heading indicators.

🧠 Heading Classification Rules:

• The first prominent, largest-font block is marked as Title.
• Large, bold blocks → H1
• Smaller but structured/indented blocks → H2/H3

🧾 JSON Output:

Identified headings are returned in a structured JSON format:

• level (Title, H1, H2, H3)
• text
• page number

✅ This avoids hardcoding and ensures generalizability + speed.

---

👤 Round 1b: Persona-Driven Document Intelligence

Round 1b builds on the structured output of Round 1a.
Goal: extract and rank sections from documents based on a persona and their "job-to-be-done".

📚 Strategy:

• Text Chunking: Use the extracted outline (Title/H1/H2/H3) to break the document into meaningful sections.
• Embedding Generation:
  • Use sentence-transformers (e.g., all-MiniLM-L6-v2)
  • Convert job description and text chunks into vector embeddings
  • Ensures semantic representation within 1GB model limit
• Relevance Scoring:
  • Compute cosine similarity between persona embedding and section embeddings.
  • Higher score = more relevant to persona's task
• Ranking:
  • Rank all sections by relevance score
  • Top-ranked ones fulfill the "Importance_rank" requirement
• Sub-section Analysis:
  • For top sections, do fine-grained text analysis
  • Provide detailed insights as per problem statement

🔄 Generalizable across documents, personas, and use cases.

---

⚙️ How to Build and Run the Solution

The solution is containerized with Docker for ease of deployment and testing.

✅ Prerequisites

• Docker installed on your system.

🧩 Models and Libraries Used PyMuPDF (fitz) For fast and accurate PDF parsing with font and layout information.

sentence-transformers For generating semantic embeddings in Round 1b (all-MiniLM-L6-v2 used).

numpy For efficient numerical computations during scoring and ranking.

🔒 All components were built with performance, generality, and container-based reproducibility in mind, ensuring a robust submission for Adobe India Hackathon 2025.

vbnet