🐉 Dragon Compressor

Neural Semantic Compression for Infinite AI Context

Resonant Pointer Architecture achieving 16:1 compression with 90%+ semantic fidelity

---

📑 Table of Contents

• What is Dragon Compressor?
• Key Features
• Performance Benchmarks
• Quick Start
• Use Cases
• Architecture Deep Dive
• Training Methodology
• Project Structure
• Running Tests & Benchmarks
• Docker Deployment
• ONNX Export
• Contributing
• Citation
• License
• Roadmap

---

🎯 What is Dragon Compressor?

Dragon Compressor solves one of the most critical problems in modern AI: memory management for long conversations and large document collections.

💡 Key Insight: Standard RAG systems store every sentence as a 384-dimensional vector. For 100,000 documents, that's 18.4 GB of RAM. Dragon Compresso compresses this to 1.15 GB while preserving 90%+ of the semantic meaning.

The Problem

• 💾 Current RAG systems waste massive memory on redundant information
• 🐌 Vector databases slow down as they scale to millions of embeddings
• 🔥 GPU memory limits force us to choose between context window and batch size

The Solution

🎯 Resonant Pointer Architecture - instead of storing all tokens, it intelligently selects and preserves only the semantic "resonance points" that carry essential meaning.

Think of it like this:

• 📚 Traditional approach: Photocopy every page of a book
• 🎯 Dragon Compressor: Extract only the key insights, quotes, and turning points

---

✨ Key Features

🚀 Extreme Compression

• 16:1 ratio (production-ready): Compress 128 tokens → 8 semantic anchors
• 64:1 ratio (experimental): Compress 128 tokens → 2 core concepts
• Maintains 90%+ cosine similarity to original embeddings

🧠 Intelligent Selection

• Resonant Pointer Mechanism: Multi-phase attention finds the most important information
• Harmonic Injection: Novel ω=6 frequency stabilization for structural coherence
• Soft Neighbor Mixing: Captures contextual information around key points

⚡ Production Ready

• Pre-trained models included (32MB)
• FastAPI server for microservice deployment
• ONNX export for C++/Rust/JavaScript integration
• Full test suite with benchmarks

🔬 Research Foundation

Built on rigorous mathematical principles:

• Hexagonal Base-6 harmonic logic (ω≈6.0)
• Teacher-Student knowledge distillation
• Multi-phase resonant pointer networks

---

📊 Performance Benchmarks

⚡ TL;DR: 16:1 compression ratio, 90%+ semantic fidelity, 100 sentences/sec, 93.8% memory savings

Memory Savings

Documents	Standard (Float32)	Dragon 1:16	Dragon 1:64	Savings
10,000	1.84 GB	0.12 GB	0.03 GB	93.5%
100,000	18.4 GB	1.15 GB	0.29 GB	93.8%
1,000,000	184 GB	11.5 GB	2.9 GB	93.8%

Quality Metrics (1:16 Ratio)

• Semantic Fidelity: 0.91 average cosine similarity
• Technical Content: 0.93 (neural networks, algorithms)
• Conversational: 0.89 (natural dialogue)
• Abstract Text: 0.90 (philosophy, literature)

Speed

• Throughput: ~100 sentences/second (RTX 5070)
• Latency: ~10ms per sentence
• Batch Processing: Scales linearly with GPU memory

---

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/Freeky7819/dragon_compressor.git
cd dragon_compressor

# Install dependencies
pip install -r requirements.txt

# Install dragon_compressor
pip install -e .

Basic Usage

from dragon.interface import Dragon

# Initialize (auto-loads pre-trained Dragon Pro 1:16)
compressor = Dragon()

# Compress a sentence
text = "Artificial intelligence is transforming how we process and store information in vector databases."

result = compressor.compress(text, ratio=16)

print(f"Original: 128 tokens")
print(f"Compressed: {result['compressed_vectors'].shape[1]} semantic anchors")
print(f"Compression: 16:1 ratio")
print(f"Positions: {result['positions']}")  # Where in text are the key points?

Output:

Original: 128 tokens
Compressed: 8 semantic anchors
Compression: 16:1 ratio
Positions: tensor([0.0234, 0.1523, 0.3125, ...])

---

📖 Use Cases

1️⃣ RAG Systems - Massive Document Libraries

from dragon.interface import Dragon
import numpy as np

dragon = Dragon()
documents = load_your_documents()  # List of 100k+ documents

# Compress entire corpus
compressed_db = []
for doc in documents:
    result = dragon.compress(doc, ratio=16)
    compressed_db.append({
        'vectors': result['compressed_vectors'],
        'positions': result['positions'],
        'original_id': doc.id
    })

# Save compressed database (93% smaller!)
np.save('compressed_knowledge_base.npy', compressed_db)

2️⃣ Long-Term AI Memory

# Compress conversation history for AI agents
conversation_history = [
    "User: What's the weather like?",
    "AI: It's sunny and 72°F.",
    "User: Should I bring an umbrella?",
    # ... 1000+ messages
]

# Compress old messages (keeps recent ones full-resolution)
old_messages = conversation_history[:-50]
compressed_memory = dragon.compress(" ".join(old_messages), ratio=16)

# AI can still "remember" key points without storing everything

3️⃣ API Microservice

# Start the FastAPI server
python API/server.py

# Client usage
import requests

response = requests.post('http://localhost:8000/compress', json={
    'text': 'Your document here...',
    'ratio': 16
})

compressed = response.json()

---

🏗️ Architecture Deep Dive

Dragon Compressor consists of three core components:

1. Harmonic Injector

Adds a decaying sinusoidal signal (ω=6.0) to embeddings, creating structural "landmarks" that survive compression.

# Hexagonal harmonic (not circular 2π)
signal = exp(-γt) × sin(6.0t + π/3)

Why ω=6? Research shows that hexagonal frequency creates more stable interference patterns in discrete latent spaces than traditional positional encodings.

2. Multi-Phase Resonant Pointer

Instead of standard attention, uses a multi-phase scanning mechanism:

• Phase 1: Broad scan for high-energy semantic regions
• Phase 2: Refined selection with LSTM memory feedback
• Confidence Gating: Dynamic weighting based on information density

3. Soft Neighbor Mixer

When a pointer selects a "key point," it also captures surrounding context using depth-wise convolutions with dilation.

# Captures ±3 token context around each selected point
Conv1D(kernel=3, padding=1) → GELU → Conv1D(kernel=3, dilation=2)

---

🔬 Training Methodology

Dragon Compressor was trained using Teacher-Student distillation:

Component	Details
Teacher Model	all-MiniLM-L6-v2 (384-dim, Hugging Face)
Dataset	WikiText-2 (2M tokens, diverse topics)
Loss Function	Cosine Similarity + Position Regularization
Optimizer	AdamW (lr=1e-4, weight_decay=0.01)
Training Time	~4 hours on RTX 5070

Validation Protocol:

• 80/20 train/test split
• Early stopping on validation cosine similarity
• Final model selected at epoch with best fidelity/compression tradeoff

---

📁 Project Structure

dragon_compressor/
├── dragon/
│   ├── __init__.py
│   ├── model.py              # Core architecture (Resonant Pointer)
│   ├── interface.py          # High-level API (Dragon class)
│   └── weights/
│       └── dragon_pro_1_16.pth  # Pre-trained model (32MB)
├── API/
│   └── server.py             # FastAPI microservice
├── demo.py                   # Interactive demo
├── eval_dragon_benchmark.py  # Full benchmark suite
├── export_onnx.py            # ONNX export for production
├── test_everything.py        # Unit + integration tests
├── requirements.txt
├── setup.py
├── Dockerfile                # Container deployment
└── README.md

---

🧪 Running Tests & Benchmarks

Full Test Suite

python test_everything.py

Tests include:

• ✅ Package import verification
• ✅ Compression tensor shape validation
• ✅ API endpoint functionality
• ✅ ONNX export compatibility

Comprehensive Benchmark

python eval_dragon_benchmark.py

Benchmark output:

• 📊 Semantic fidelity across diverse text types
• 🔍 Pointer interpretability ("X-ray" visualization)
• 💾 Memory usage calculations
• ⚡ Throughput measurements

Interactive Demo

python demo.py

---

🐳 Docker Deployment

# Build image
docker build -t dragon_compressor:latest .

# Run container
docker run -p 8000:8000 dragon_compressor:latest

# Test API
curl -X POST http://localhost:8000/compress \
  -H "Content-Type: application/json" \
  -d '{"text": "Test compression", "ratio": 16}'

---

🔧 ONNX Export (C++/Rust Integration)

python export_onnx.py

This generates dragon_1_16.onnx which can be loaded in:

• C++: ONNX Runtime
• Rust: tract or onnxruntime-rs
• JavaScript: onnxruntime-web
• C#: ML.NET

Example (C++):

#include <onnxruntime/core/session/onnxruntime_cxx_api.h>

Ort::Env env;
Ort::Session session(env, "dragon_1_16.onnx", Ort::SessionOptions());

// Run inference
auto output = session.Run(...);

---

🤝 Contributing

We welcome contributions! Here's how you can help:

Areas of Interest

• 🎯 Adaptive Compression: Dynamic ratio selection based on content complexity
• 🌐 Multilingual Support: Extend beyond English (currently optimized for English)
• 📊 Benchmark Datasets: Test on domain-specific corpora (medical, legal, code)
• ⚡ Performance: CUDA kernel optimizations, quantization

Development Setup

# Fork & clone
git clone https://github.com/Freeky7819/dragon_compressor.git

# Create feature branch
git checkout -b feature/your-feature-name

# Install dev dependencies
pip install -r requirements-dev.txt

# Run tests before committing
python test_everything.py

# Submit PR
git push origin feature/your-feature-name

---

📚 Citation

If you use Dragon Compressior in your research, please cite:

@software{dragon_compressor 2025,
  title={Dragon Compressor Resonant Semantic Compression for Infinite AI Context},
  author={Žakelj, Damjan},
  year={2024},
  url={https://github.com/Freeky7819/dragon_compressor},
  note={Neural architecture achieving 16:1 compression with 90\%+ semantic fidelity}
}

Related Research:

• Hexagonal Harmonic Stabilization in Latent Spaces (ω=6 phenomenon)
• Multi-Phase Resonant Pointer Networks
• Teacher-Student Distillation for Semantic Compression

---

📄 License

This project is licensed under the MIT License - see LICENSE file for details.

TL;DR: Free to use, modify, and distribute. Commercial use allowed. Just keep the license notice.

---

🙏 Acknowledgments

• Sentence-Transformers: For the excellent all-MiniLM-L6-v2 model
• PyTorch Team: For the incredible deep learning framework
• FastAPI: For the elegant API framework
• Community: For testing, feedback, and contributions

---

📞 Contact & Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Email: [email protected]

---

🎯 Roadmap

v1.1 (Q1 2025)

• Adaptive ratio selection based on content entropy
• Multi-GPU training support
• Quantized models (INT8, FP16)

v1.2 (Q2 2025)

• Multilingual models (50+ languages)
• Online learning / incremental compression
• Integration with LangChain & LlamaIndex

v2.0 (Q3 2025)

• Hierarchical compression (compress compressed vectors)
• Real-time streaming compression
• Hardware acceleration (TPU, Apple Silicon)

---

🐉 Built with precision, passion, and mathematical poetry 💙

Star ⭐ this repo if Dragon Compressor helps your project!

⬆ Back to Top