CosineFusion 🚀

Overview

CosineFusion is a hardware-accelerated, high-performance cosine similarity engine that seamlessly bridges C++17 and Python using pybind11. By executing intense linear algebra routines directly on compiled binaries, CosineFusion bypasses the Python interpreter's performance bottlenecks, delivering bare-metal speed for matrix matching, AI embeddings, and content recommendation pipelines.

🔥 Key Features

• Bare-Metal Performance: High-speed C++ back-end optimized for vector and matrix calculations.
• Universal2 Multi-Chip Support: Natively compiled to support both Intel (x86_64) and Apple Silicon M-Series (arm64) Mac architectures out of the box with zero manual configuration.
• Seamless Python Binding: Exposes highly optimized native bindings directly to Python workflows using pybind11.
• Production-Grade Architecture: Restructured using the standard Python src-layout to ensure strict dependency separation, clean imports, and absolute deployment stability.

🎯 Use Cases & Applications

• Enterprise Recommender Systems: Instantaneous user-to-item feature matrix matching for e-commerce or media content streaming.
• Personalization Engines: Real-time scoring based on multi-dimensional user preference vectors.
• Fast Similarity Search: Sub-millisecond similarity scoring for AI/ML feature embedding matching.

💻 Installation

Option 1: Direct from PyPI (Recommended Production Method)

Install the pre-compiled, optimized universal wheels directly from the global package registry:

pip install cosinefusion

Option 2: Local Development Setup

If working from source, clone the repository and build the native extensions locally:

git clone [https://github.com/RoughneckCoder/CosineFusion.git](https://github.com/RoughneckCoder/CosineFusion.git)
cd CosineFusion
pip install pybind11
pip install .

⚡ Usage Example

import numpy as np
import core_init

# Example item features
items = np.array([
    [0, 1, 0, 0, 1], # Tea
    [0, 1, 0, 0, 1], # Coffee
    [1, 0, 1, 1, 0], # Jaffa Cake
    [1, 0, 1, 0, 0], # Biscuit
    [1, 0, 1, 1, 1], # Chocolate Bar
    [0, 1, 0, 0, 1], # Espresso
])

item_names = ["Tea", "Coffee", "Jaffa Cake", "Biscuit", "Chocolate Bar", "Espresso"]

# User likes sweetness and chocolate
user = np.array([[0, 0, 0, 1, 0]]) 

# Executes natively on the underlying hardware (Intel or Apple Silicon)
res = core_init.cosine_similarity(user, items)
sim = res["similarity_matrix"]

top_indices = np.argsort(sim[0])[::-1]
print("Top recommendations for user:")
for i in top_indices:
    print(f"{item_names[i]}: {sim[0][i]:.2f}")

Sample Output:

Top recommendations for user:
Chocolate Bar: 0.89
Jaffa Cake: 0.75     <- Accurately indicates structural sweet-matrix alignment
Biscuit: 0.65
Tea: 0.45
Coffee: 0.45
Espresso: 0.43

🏗️ Hardware Architecture & Multi-Chip Targeting

To maximize cross-platform utility, the build pipeline leverages fat multi-architecture binary compilation. When deploying the package via pip install, the distribution system automatically delivers a native Universal2 binary.

This ensures that the underlying C++ extension automatically executes instructions matched perfectly to the host machine's physical CPU:

• Apple Silicon (M1/M2/M3/M4 chips): Executes native ARM64 instructions.
• Intel Processors: Executes native x86_64 instructions.

🧪 Testing

The repository includes an automated verification suite to guarantee mathematical accuracy and cross-platform compatibility. Run tests using pytest:

python -m pytest -v

📁 Project Structure

CosineFusion/
 ├── src/
 │    ├── cpp/
 │    │    └── core_init.cpp     # High-performance C++ core engine
 │    └── python/
 │         └── core_demo.py     # Functional verification demo
 ├── tests/
 │    └── test_bridge.py        # Automated PyTest matrix suite
 ├── setup.py                   # Pybind11 / Universal2 compilation configuration
 ├── pyproject.toml             # Standard modern build-system requirements
 ├── README.md                  # Project documentation
 ├── LICENSE                    # MIT License open-source file
 └── requirements.txt           # Environment dependencies

📄 License

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

Author: Sam Chaudry

GitHub: RoughneckCoder