VectorCore: Low-Level Vector Search & HNSW Indexing Engine

VectorCore is a lightweight, zero-dependency Approximate Nearest Neighbor (ANN) vector search engine built from scratch in Python and NumPy. It implements SIMD-friendly vector distance metrics, an exact brute-force baseline index, and a Hierarchical Navigable Small World (HNSW) graph index with binary disk serialization.

---

Key Features

• Vectorized Metric Kernels: Optimized Euclidean ($L_2$) and Cosine distance metric routines.
• HNSW Graph Index: Fast Approximate Nearest Neighbor (ANN) greedy graph traversal with configurable ef_construction, M, and ef_search beam width parameters.
• Exact Flat Index: Exhaustive linear-scan baseline providing 100% ground-truth recall validation.
• Binary Serialization: Zero-copy disk persistence protocol (.vcore) preserving graph topologies and vector payload matrices.
• Production Distribution: Published on PyPI as vectorcore-ann with automated CI/CD releases.

---

Benchmark Results

Evaluated on 5,000 vectors (128 dimensions) queried with 100 randomized vectors at $k=10$:

Index Type	Build Time	Avg Latency	Throughput (QPS)	Recall@10
Flat (Brute-Force)	0.000s	1.298 ms	770.3 queries/s	100.0%
HNSW (Graph ANN)	6.830s	0.533 ms	1876.5 queries/s	61.3%

• Performance Gain: 2.44x faster search latency over brute-force linear scanning.

---

Installation

Via PyPI (Recommended)

pip install vectorcore-ann