ZeroVector.Core 1.0.0

📐 ZeroVector: Sovereign Embedded Vector Database & SIMD Similarity Engine

ZeroVector is an ultra-high-throughput, zero-allocation embedded Vector Database and SIMD similarity metric engine engineered in 100% pure C# for .NET. It resides in Tier 2 (Transport & Storage) of the ZeroPlatform ecosystem.

⚡ Key Capabilities

• Hardware-Accelerated SIMD Metrics (AVX2 / FMA):
  • DotProduct: Simultaneous 8-wide float multiply-add evaluation.
  • CosineSimilarity: Single-pass combined dot product and norm calculation ($\le 1.0$).
  • EuclideanDistance & EuclideanDistanceSquared: L2 metric for spatial embeddings.
  • ManhattanDistance: L1 metric for sparse features.
  • HammingDistance: 64-bit hardware POPCNT for binary bitstrings (ORB/BRIEF).
• Vector Indices:
  • FlatVectorIndex: Contiguous memory, exact brute-force Top-K search with multi-core parallel partitioning for up to millions of embeddings.
  • HnswVectorIndex: Hierarchical Navigable Small World (HNSW) graph index providing sub-millisecond $O(\log N)$ approximate nearest neighbor (ANN) search with configurable $M$, $efConstruction$, and $efSearch$.
• Storage & Persistence:
  • Compact zero-overhead binary index serialization and deserialization (VectorIndexFile).
• Zero External Dependencies & Multi-Targeting:
  • Pure C# implementation compatible with .NET 8.0+, .NET Framework 4.6.2+, and .NET Standard 2.0.

🚀 Quick Start

1. Flat Contiguous Index (Exact Search)

using ZeroVector.Core.Indices;
using ZeroVector.Core.Metrics;

// Initialize index for 384-dimensional embeddings
var index = new FlatVectorIndex(dimension: 384, initialCapacity: 1024);

// Add embeddings
ReadOnlySpan<float> embedding = stackalloc float[384];
// ... populate embedding ...
index.Add(id: 1, embedding);

// Query Top-K nearest neighbors via Cosine Similarity
ReadOnlySpan<float> query = stackalloc float[384];
VectorSearchResult[] results = index.SearchTopK(query, k: 5, VectorMetricType.Cosine);

foreach (var r in results)
{
    Console.WriteLine($"Matched ID: {r.Id}, Similarity Score: {r.Score:F4}");
}

2. HNSW Graph Index (Fast Approximate Search)

using ZeroVector.Core.Indices;
using ZeroVector.Core.Metrics;

// Initialize HNSW index
var hnsw = new HnswVectorIndex(
    dimension: 128,
    m: 16,
    efConstruction: 100,
    efSearch: 50,
    metric: VectorMetricType.Cosine
);

// Populate index
for (int i = 0; i < 10000; i++)
{
    float[] vec = GenerateVector(128);
    hnsw.Add(id: i, vec);
}

// Sub-millisecond ANN search
VectorSearchResult[] matches = hnsw.SearchTopK(queryVec, k: 10);

🏛 Architectural Placement in ZeroPlatform

ZeroVector resides strictly within Tier 2 (Transport & Storage):

• Upstream consumers: ZeroInference (L3), ZeroGraphics (L4), ZeroAgent (L5), and ZeroPipeline (L5).
• Downstream dependencies: ZeroPrimitives (L0) and pure .NET primitives.

📄 License

Architected and developed by Phong Võ (kzxl). Released under the MIT License.