Torpe Hitachi Classifier

Max Winning Project: Gemini-Centric Document Classification System

A comprehensive, enterprise-grade document classification system powered by Google's Gemini 2.0 Flash, featuring RAG (Retrieval Augmented Generation), CAG (Context Augmented Generation), Solana blockchain audit trails.

🌟 Features

Phase 1: Foundation & Policy RAG

• ✅ Policy Knowledge Base: Comprehensive category definitions, PII patterns, and SME-validated examples
• ✅ Multi-Modal Document Processing: PDF parsing with OCR for text and images
• ✅ Citation Mapping: Precise source location tracking with bounding boxes
• ✅ Gemini File Search Store: RAG-based policy grounding

Phase 2: Core AI Engine with RAG/CAG

• ✅ Dynamic Prompt Tree: Sequential classification flow (UNSAFE → CONFIDENTIAL → SENSITIVE → PUBLIC)
• ✅ RAG + CAG Grounding: Policy knowledge base + cached document content
• ✅ Structured JSON Output: Category, confidence, reasoning, and citations
• ✅ Dual-Layer Validation: Consensus-based auto-approval (90%+ confidence threshold)

Phase 3: Auditability, UX & Compliance

• ✅ Solana Blockchain: Immutable audit trails on Solana devnet
• ✅ SQLite Audit Logs: Complete classification history and HITL reviews
• ✅ Web UI: Flask-based interface with HITL feedback loop

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│                    Document Upload (PDF)                     │
└─────────────────┬───────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────┐
│  Document Processing (PyMuPDF + OCR + Citation Mapping)     │
└─────────────────┬───────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────┐
│         Gemini Classifier (RAG + CAG Pipeline)              │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐     │
│  │ Policy RAG   │  │ Cached Doc   │  │ Dual Layer   │     │
│  │ (File Search)│ +│ (CAG)        │ +│ Validation   │     │
│  └──────────────┘  └──────────────┘  └──────────────┘     │
└─────────────────┬───────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────┐
│              Classification Result + Metadata                │
└─────────┬───────────┬───────────┬───────────────────────────┘
          │           │           │
          ▼           ▼           ▼
┌──────────────┐ ┌──────────────┐
│   Solana     │ │SQLite Audit  │
│ Blockchain   │ │   Logger     │
└──────────────┘ └──────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────┐
│        Web UI (Dashboard + HITL Review Queue)               │
└─────────────────────────────────────────────────────────────┘

📋 Requirements

• Python 3.9+
• Tesseract OCR
• API Keys:
  • Google Gemini API
  • Solana Devnet access

🚀 Installation

1. Clone/Navigate to Project

cd gemini-classifier

2. Install System Dependencies

macOS:

brew install tesseract

Ubuntu/Debian:

sudo apt-get update
sudo apt-get install tesseract-ocr

Windows: Download and install Tesseract from: https://github.com/UB-Mannheim/tesseract/wiki

3. Create Virtual Environment

python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

4. Install Python Dependencies

pip install -r requirements.txt

5. Configure Environment Variables

The .env file is already configured with your API keys:

GEMINI_API_KEY=AIzaSyA5CRA7vt5rLIVzrW9mTFOTMtFCasEhxlo
SOLANA_CLUSTER_URL=https://api.devnet.solana.com

Note: In production, use environment variables or secure secret management instead of committing API keys.

🎯 Usage

Start the Web Application

python main.py

The application will be available at:

• Main Upload: http://localhost:5000
• Dashboard: http://localhost:5000/dashboard
• HITL Queue: http://localhost:5000/hitl/queue

Classify a Document

1. Navigate to http://localhost:5000
2. Upload a PDF file (drag-and-drop or click to browse)
3. Wait for processing (typically 5-15 seconds)
4. Review the classification result with:
   • Category (UNSAFE/CONFIDENTIAL/SENSITIVE/PUBLIC)
   • Confidence score
   • Reasoning and citations
   • Blockchain audit hash

HITL Review Process

1. Navigate to HITL Queue (http://localhost:5000/hitl/queue)
2. Click "Review Document" on any pending classification
3. Verify or correct the classification
4. Add reviewer notes
5. Submit review

Important: Corrected classifications are automatically added to the RAG knowledge base as new few-shot examples, improving future accuracy.

📊 Classification Categories

1. UNSAFE (Priority 1)

• Harmful, violent, or threatening content
• Illegal activity instructions
• Malware or security exploits
• Action: Immediate rejection and escalation

2. CONFIDENTIAL (Priority 2)

• Trade secrets and proprietary algorithms
• Financial records (with SSN, credit cards)
• Legal documents (attorney-client privilege)
• M&A plans, executive compensation
• Source code and IP
• PII: SSN, credit cards, bank accounts, medical records, passports

3. SENSITIVE (Priority 3)

• Internal memos and communications
• Employee directories
• Draft documents
• Internal project plans
• Non-executive budgets
• PII: Emails, phone numbers, addresses, employee IDs

4. PUBLIC (Priority 4)

• Published marketing materials
• Public website content
• Press releases
• Open-source code
• Public documentation

🔧 API Endpoints

Upload and Classify

POST /upload
Content-Type: multipart/form-data

Response: {
  "document_id": "DOC_abc123...",
  "classification": "CONFIDENTIAL",
  "confidence": 0.95,
  "reasoning": "...",
  "citation": "...",
  "blockchain": {...},
  "audio_available": true
}

Get Statistics

GET /api/statistics

Response: {
  "total_classifications": 42,
  "auto_approval_rate": 85.5,
  "avg_processing_time": 8.3,
  "by_category": {...}
}

Get All Classifications

GET /api/classifications?limit=100&offset=0

Get Specific Classification

GET /api/classification/<document_id>

Submit HITL Review

POST /hitl/submit

## 🧪 Testing

### Test with Sample Documents

Create test PDFs with different content types:

**Confidential Example:**

CONFIDENTIAL - Board Meeting Minutes Acquisition Target: TechCorp Offer: $500M Employee Data: John Smith - SSN: 123-45-6789 Credit Card: 4532-1234-5678-9010

**Public Example:**

FOR IMMEDIATE RELEASE Product Launch Announcement Contact: press@company.com

### Verify System Components

1. **RAG Policy Upload**: Check console for "Policy uploaded successfully"
2. **Classification**: Verify JSON output with category, confidence, reasoning
3. **Blockchain**: Check for transaction hash (may be simulated if devnet is down)
4. **Database**: SQLite file at `data/audit_logs.db`

## 📁 Project Structure

gemini-classifier/ ├── main.py # Main entry point ├── requirements.txt # Python dependencies ├── .env # Environment variables (API keys) ├── README.md # This file ├── src/ │ ├── config.py # Configuration management │ ├── audit_logger.py # SQLite audit logging │ ├── processing/ │ │ ├── init.py │ │ └── document_processor.py # PDF/OCR processing │ ├── classification/ │ │ ├── init.py │ │ ├── policy_rag.py # RAG knowledge base │ │ └── classifier.py # Core AI classifier │ ├── blockchain/ │ │ ├── init.py │ │ └── solana_audit.py # Solana integration │ └── ui/ │ ├── init.py │ └── app.py # Flask web application ├── policies/ │ ├── categories.json # Category definitions │ ├── pii_patterns.json # PII detection patterns │ └── few_shot_examples.json # SME-validated examples ├── templates/ │ ├── base.html │ ├── index.html # Upload page │ ├── dashboard.html # Statistics dashboard │ ├── hitl_queue.html # Review queue │ └── hitl_review.html # Review detail page └── data/ ├── uploads/ # Uploaded PDFs ├── cache/ # Cached content & audio ├── audit_logs/ # Log files └── audit_logs.db # SQLite database

## 🎓 Key Technologies

| Component | Technology | Purpose |
|-----------|-----------|---------|
| AI Model | Gemini 2.0 Flash | Fast, high-quality classification |
| RAG | Gemini File Search | Policy knowledge grounding |
| CAG | Gemini Caching API | Document context optimization |
| Blockchain | Solana (Devnet) | Immutable audit trails |
| Database | SQLite | Local audit logging |
| Web Framework | Flask | REST API & web UI |
| OCR | Tesseract + PyMuPDF | Multi-modal document processing |

## 🔐 Security Considerations

1. **API Keys**: Never commit API keys to version control. Use environment variables.
2. **PII Detection**: High-risk PII triggers CONFIDENTIAL classification.
3. **Audit Trail**: All decisions are logged to SQLite and Solana blockchain.
4. **HITL Review**: Human oversight for low-confidence or mismatched validations.
5. **Safety Checks**: UNSAFE content is detected first and rejected immediately.

## 📈 Performance Metrics

- **Processing Speed**: ~5-15 seconds per document (depends on page count)
- **Auto-Approval Rate**: Target 85%+ with dual validation
- **Confidence Threshold**: 90% for auto-approval

## 🐛 Troubleshooting

### "Tesseract not found"
Install Tesseract OCR (see Installation section)

### "File processing failed"
Check that the PDF is not corrupted or password-protected

### "Blockchain recording error"
The system will create a simulated transaction hash if Solana devnet is unavailable. This is normal for demo purposes.

### Gemini API errors
- Check API key validity
- Verify quota/billing is enabled
- Ensure Gemini 2.0 Flash access is enabled

## 🤝 HITL Feedback Loop

The system implements a continuous improvement cycle:

1. Document is classified by AI
2. If confidence < 90% or dual validation mismatch → HITL queue
3. SME reviews and corrects classification
4. Correction is added to `policies/few_shot_examples.json`
5. Policy RAG is updated automatically
6. Future similar documents benefit from the correction

## 📝 License

This is a demonstration project for educational purposes.

## 🙏 Acknowledgments

- **Google Gemini**: Advanced AI classification engine
- **Solana**: Blockchain infrastructure for audit trails
- **Tesseract OCR**: Open-source OCR engine

## 📞 Support

For issues or questions about this implementation, please review:
1. This README
2. The code comments (extensively documented)
3. The policy JSON files in `policies/` directory

---

**Built with ❤️ using Gemini 2.0 Flash and Solana**