QEPM-1K: Sub-Linear Knowledge Retrieval Code

✅ Independent Verification Results

Status: Paper claims fully verified (December 2025)

Metric	Paper Claim	Verified Result	Status
Accuracy	100%	100.0%	✅ VERIFIED
Retrieval Time	0.88ms	0.034ms	✅ EXCEEDED (25.9× faster)
Exact Bucket Hits	93%	100%	✅ EXCEEDED

See verification/VERIFIED_RESULTS_100PCT.md for complete details.

---

Published Research

Paper: Horn, J.P. (2025). Sub-Linear Knowledge Retrieval via Quantum-Inspired Hyperdimensional Folded Space. Zenodo. https://doi.org/10.5281/zenodo.17848904

---

📦 What's Included

This repository contains the complete implementation of the QEPM-1K knowledge retrieval system described in the paper.

Core Files:

1. build_qepm_1k.py - Builds 1,100 Q&A knowledge base
2. test_1k_folded_space.py - Tests with 4D folded space indexing
3. quantum_hdc_encoder_optimized.py - 10,000D HDC encoder
4. quantum_inference_optimized_v2.py - Pattern inference engine
5. entangle_storage_fixed.py - PHOTON storage format
6. quantum_training_ultra.py - Training system (203 patterns/sec)
7. hdc_utils.py - Core HDC utilities
8. deployment_guide.py - Technical documentation

Documentation:

9. README.md - This file
10. EXECUTIVE_SUMMARY.md - Performance overview

---

🚀 Quick Start

Prerequisites

# Python 3.10+
pip install numpy numba

Build Knowledge Base

# Build 1,100 Q&A knowledge base (~24 seconds)
python build_qepm_1k.py

Output: scaling/qepm_knowledge_1k_output/qepm_knowledge_1k/

Test Retrieval

# Test with 4D folded space indexing
python test_1k_folded_space.py

Expected Results:

• Accuracy: 100% (15/15 queries)
• Average speed: 0.88ms
• Exact bucket hits: 93%

---

📊 Reproducing Paper Results

Main Result (Table 1)

python test_1k_folded_space.py

Expected output:

Overall Performance:
- Accuracy: 100.0% (15/15)
- Average time: 0.88ms
- Median time: 0.78ms
- Throughput: 1,140 queries/second

Folded Space Strategy:
- Exact bucket: 93% (14/15)
- 1-hop neighbors: 7% (1/15)
- Full search: 0% (0/15)

Hardware Requirements

Minimum:

• CPU: Intel Celeron N4020 @ 1.1GHz (or equivalent)
• RAM: 8 GB
• Storage: 1 GB available
• OS: Windows/Linux/macOS

Note: No GPU required!

---

📁 File Structure

QEPM-1K-Code/
├── README.md                              # This file
├── EXECUTIVE_SUMMARY.md                   # Performance overview
│
├── build_qepm_1k.py                       # Build knowledge base
├── test_1k_folded_space.py               # Test & validate
│
├── quantum_hdc_encoder_optimized.py      # HDC encoder (10,000D)
├── quantum_inference_optimized_v2.py     # Inference engine
├── entangle_storage_fixed.py             # Storage system
├── quantum_training_ultra.py             # Training (203 pat/sec)
├── hdc_utils.py                          # Core utilities
│
└── deployment_guide.py                    # Technical docs

---

🔧 Usage Examples

Example 1: Build Custom Knowledge Base

from build_qepm_1k import build_knowledge_base

# Create your own Q&A pairs
data = [
    ("What is AI?", "Artificial Intelligence is..."),
    ("What is ML?", "Machine Learning is..."),
    # ... add more
]

# Build knowledge base
build_knowledge_base(
    data=data,
    output_path="my_knowledge_base/",
    model_dim=2048
)

Example 2: Query Knowledge Base

from test_1k_folded_space import FoldedSpaceKnowledgeQEPM

# Load knowledge base
kb = FoldedSpaceKnowledgeQEPM(
    model_path="scaling/qepm_knowledge_1k_output/qepm_knowledge_1k/"
)

# Query
answer, confidence, strategy = kb.query("What is machine learning?")

print(f"Answer: {answer}")
print(f"Confidence: {confidence:.2%}")
print(f"Strategy: {strategy}")

---

📈 Performance Benchmarks

On Intel Celeron N4020 @ 1.1GHz:

Metric	Value
Knowledge pairs	1,100
Accuracy	100% (15/15 test queries)
Average speed	0.88ms
Median speed	0.78ms
Min speed	0.59ms
Max speed	1.30ms
Throughput	1,140 queries/second
Exact bucket hits	93% (O(1) retrieval)

Comparison:

• vs. Exhaustive search: 162× faster
• vs. 80-pair baseline: 13× faster with 13.75× more data

---

🧪 Testing

Run all tests:

# Test HDC encoder
python quantum_hdc_encoder_optimized.py

# Test inference engine
python quantum_inference_optimized_v2.py

# Test storage
python entangle_storage_fixed.py

# Test complete system
python test_1k_folded_space.py

---

🛠️ Customization

Change Knowledge Base Size

Edit build_qepm_1k.py:

# Change from 1,100 to your desired size
PAIRS_PER_DOMAIN = {
    'ml_ai': 100,           # Change these numbers
    'computer_science': 100,
    # ...
}

Change Folded Space Resolution

Edit test_1k_folded_space.py:

# Change from 7×7×7×7 to different size
self.grid_size = 7  # Try 5, 9, 11, etc.

Recommendation:

• Small KB (<500): grid_size = 5
• Medium KB (500-2000): grid_size = 7
• Large KB (2000+): grid_size = 9

---

📖 Citation

If you use this code, please cite:

@misc{horn2025sublinear,
  title={Sub-Linear Knowledge Retrieval via Quantum-Inspired Hyperdimensional Folded Space},
  author={Horn, Jared Paul},
  year={2025},
  month={December},
  publisher={Zenodo},
  doi={10.5281/zenodo.17848904},
  url={https://doi.org/10.5281/zenodo.17848904}
}

---

📧 Contact

Jared Paul Horn
Email: [email protected]
DOI: 10.5281/zenodo.17848904

---

📄 License

This code is provided for research and educational purposes.

You are free to:

• Use the code for research
• Modify and adapt
• Include in your projects

Please:

• Cite the paper if you use this code
• Share improvements back to the community

---

🙏 Acknowledgments

This work was completed on consumer hardware (Intel Celeron N4020, 12GB RAM) to demonstrate that advanced AI research doesn't require expensive GPUs.

---

🐛 Known Issues

None currently. If you find bugs, please contact: [email protected]

---

📝 Changelog

v1.0 (December 2025)

• Initial release
• 1,100 Q&A knowledge base
• 0.88ms average retrieval
• 100% accuracy on test set

---

⚡ Quick Reference

Build knowledge base:

python build_qepm_1k.py

Test system:

python test_1k_folded_space.py

Expected results:

• 100% accuracy
• 0.88ms average speed
• 93% exact bucket hits

Hardware:

• Consumer laptop (no GPU)
• ~$200 cost
• 12GB RAM sufficient

Performance:

• 162× faster than exhaustive search
• 13× faster than 80-pair baseline
• 1,140 queries/second throughput

---

For questions or issues, contact: [email protected]

---

⭐ Found This Useful?

If this research helped you or you'd like to support open science, please star this repository!

Questions? Open an issue or reach out: [email protected]