HyperSearch

What I Built

HyperSearch is a C++ semantic search engine built from first principles. It converts text into 384-dimensional embeddings and performs similarity-based retrieval across collections of up to 100,000 vectors.

Rather than treating vector search as a black box, I built and optimized the core retrieval pipeline myself, using the project to investigate how algorithms, memory layout, multithreading, and CPU-level vectorization affect real-world search performance.

How It Works

The initial implementation used brute-force similarity search. Every query was compared against the stored vectors, providing a simple and measurable baseline before optimization.

Performance Engineering

I established a reproducible benchmark across collections ranging from 1,000 to 100,000 vectors. The initial brute-force implementation reached approximately 1.10 seconds of median latency at 100K vectors and 4.19 queries per second.

I then optimized the retrieval pipeline through multiple layers: OpenMP multithreading to distribute search work, AVX2/FMA SIMD instructions to accelerate numerical computation, and memory-layout improvements to make vector access more efficient.

Results

Engineering Lessons

The project taught me that performance optimization starts with measurement rather than assumptions. I established a baseline, identified the expensive parts of the retrieval pipeline, applied one class of optimization at a time, and measured the result.

HyperSearch ultimately became less about building a search feature and more about understanding the interaction between algorithms, memory, parallelism, and CPU architecture.