🤖 DeckBot - PDF Pitch Deck Metadata Extractor

A sophisticated RAG-optimized metadata extraction system for pitch deck PDFs. Processes pitch decks to generate structured metadata at both deck-level (global context) and slide-level (detailed content) for high-performance retrieval systems.

🎯 Features

• Multi-Level Chunking: Extracts both deck-level and slide-level metadata
• Industry Classification: 30+ industry categories with ENUM enforcement
• Slide Layout Detection: Automatic classification of slide types (Title, Problem, Solution, etc.)
• Gemini 2.0 Flash: Powered by Google's latest AI model for accurate extraction
• Batch Processing: Process entire folders of PDFs automatically
• Validation & Statistics: Built-in validation and detailed analytics
• RAG-Optimized: Structured for vector database storage and semantic search

📊 Data Structure

Deck-Level Metadata (Global Context)

{
  deck_industry: DeckIndustry,      // ENUM: Fintech, SaaS, etc.
  company_name: string,              // Extracted company name
  executive_summary: string,         // 2-3 sentence deck summary
  total_pages: number               // Total slide count
}

Slide-Level Metadata (Detailed Content)

{
  filename: string,                  // Source PDF filename
  slide_number: number,              // Page index
  slide_content: string,             // Full text extraction
  slide_summary: string,             // 1-2 sentence summary
  keywords: string[],                // 3-5 core keywords
  slide_layout: SlideLayout         // ENUM: Title, Problem, etc.
}

🚀 Quick Start

Prerequisites

• Node.js 18+ and npm
• Gemini API key (Get one here)
• poppler-utils for PDF processing:

  # macOS
  brew install poppler
  
  # Ubuntu/Debian
  sudo apt-get install poppler-utils
  
  # Windows
  # Download from: https://github.com/oschwartz10612/poppler-windows/releases

Installation

1. Clone and install dependencies:

   npm install

2. Configure environment:

   cp .env.example .env
   # Edit .env and add your GEMINI_API_KEY

3. Add PDFs to process:

   # Place your pitch deck PDFs in the pdf/ directory

Usage

Process all PDFs:

npm run dev

Build and run:

npm run build
npm start

📁 Project Structure

DeckBot/
├── src/
│   ├── types.ts              # TypeScript type definitions & ENUMs
│   ├── pdf-processor.ts      # PDF text extraction & image conversion
│   ├── gemini-client.ts      # Gemini API integration
│   ├── metadata-extractor.ts # Main extraction orchestration
│   └── index.ts              # CLI entry point
├── pdf/                      # Input PDFs (place files here)
├── output/                   # Generated JSON metadata files
├── images/                   # Converted slide images (temporary)
├── package.json
├── tsconfig.json
└── README.md

🔧 Configuration

Edit src/index.ts to customize:

const config: ProcessingConfig = {
  pdfDirectory: path.join(projectRoot, 'pdf'),
  outputDirectory: path.join(projectRoot, 'output'),
  imageDirectory: path.join(projectRoot, 'images'),
  geminiApiKey: apiKey,
  maxConcurrentRequests: 3  // Adjust for rate limiting
};

📋 Industry ENUMs

The system supports 30+ industry classifications:

• Fintech, E-commerce, SaaS
• HealthTech, EdTech, PropTech
• Marketplace, Enterprise Software
• Consumer App, Biotech, CleanTech
• AI/ML, Cybersecurity, Web3/Crypto
• Social Media, Gaming, Hardware
• Marketing/AdTech, HR/Recruiting
• ... and more (see src/types.ts)

📋 Slide Layout Types

Automatic detection of common pitch deck slide types:

• Title, Problem, Solution
• Product, Traction, Market
• Business Model, Competition
• Team, Financials, Roadmap
• Ask, Contact, Appendix

📊 Output Format

Each PDF generates a JSON file with complete metadata:

{
  "deck_metadata": {
    "deck_industry": "Fintech",
    "company_name": "Example Corp",
    "executive_summary": "A neobank focused on...",
    "total_pages": 25
  },
  "slides": [
    {
      "filename": "pitch_deck.pdf",
      "slide_number": 1,
      "slide_content": "Complete extracted text...",
      "slide_summary": "Title slide introducing...",
      "keywords": ["fintech", "banking", "mobile"],
      "slide_layout": "Title"
    }
    // ... more slides
  ]
}

🎯 RAG Integration Strategy

Storage Pattern

1. Vector DB: Store each slide as a separate document
2. Metadata: Attach both slide-level AND deck-level metadata
3. Embeddings: Generate from slide_summary + slide_content

Retrieval Pattern

1. Semantic Search: Query against slide embeddings
2. Metadata Filtering: Filter by deck_industry, slide_layout
3. Context Assembly: Include deck-level context with retrieved slides

Example Query Flow

// 1. Semantic search
const results = await vectorDB.search(query, {
  filter: { deck_industry: 'Fintech' },
  limit: 5
});

// 2. Assemble context
const context = results.map(slide => ({
  global: slide.metadata.deck_metadata,
  local: slide.metadata.slide_content
}));

⚙️ Processing Pipeline

1. PDF Text Extraction: Extract complete text using pdf-parse
2. Deck Summarization: Generate global context with Gemini
3. Image Conversion: Convert each page to PNG images
4. Slide Analysis: Process each image with Gemini vision
5. Metadata Generation: Extract structured data for each slide
6. Validation: Verify completeness and consistency
7. JSON Export: Save structured output files

🚦 Rate Limiting

The system includes built-in delays between API calls:

• 1 second delay between slide processing
• Configurable maxConcurrentRequests
• Adjust in src/metadata-extractor.ts as needed

🐛 Troubleshooting

"poppler not found"

Install poppler-utils (see Prerequisites)

"GEMINI_API_KEY not found"

Ensure .env file exists with valid API key

Rate limiting errors

Increase delay in metadata-extractor.ts:

await this.geminiClient.delay(2000); // 2 seconds

Image conversion fails

Check PDF permissions and poppler installation

📈 Performance

• Processing Speed: ~5-10 seconds per slide (depends on API)
• Batch Processing: Supports unlimited PDFs sequentially
• Memory Usage: ~100-500MB per PDF
• API Costs: ~$0.01-0.05 per deck (Gemini 2.0 Flash pricing)

🔐 Security

• API keys stored in .env (never commit!)
• All processing happens locally
• Images stored temporarily (auto-cleanup optional)
• No external data transmission except to Gemini API

📝 License

MIT

🤝 Contributing

Contributions welcome! Please open an issue or PR.

🙏 Acknowledgments

Built with:

• Google Generative AI - Gemini 2.0 Flash
• pdf-parse - PDF text extraction
• pdf-poppler - PDF to image conversion
• sharp - Image processing

---

Made with ❤️ for RAG systems and pitch deck analysis