Glacier.Vector 1.0.7

Glacier.Vector

📖 Read the Deep-Dive: I built a zero-dependency C# vector database that saturates DDR5 RAM bandwidth.

========================================================================================================
  GLACIER.VECTOR BARE-METAL GPU THROUGHPUT BENCHMARK (NVIDIA RTX 4060 vs PYTHON FAISS)
========================================================================================================
  Python FAISS (CPU Brute Force)     :  ~120 ms
  Glacier.Vector (Bare-Metal GPU)    :  2.57 ms (622,700,000 vectors/sec)
--------------------------------------------------------------------------------------------------------
  🏆 SPEEDUP                         :  46x FASTER THAN PYTHON FAISS / 10.1x FASTER THAN AVX-512
  ⚡ HARDWARE ARCHITECTURE           :  Pure C# Direct P/Invoke (Zero CUDA Toolkit / Zero C++ DLLs)
========================================================================================================

Glacier.Vector is a high-performance, SIMD-accelerated vector search engine and index for .NET 10. Built for speed and efficiency, it provides a zero-copy architecture for storing and querying high-dimensional embeddings, commonly used in LLM-powered applications and semantic search.

Key Features

• 🚀 SIMD-Accelerated Search: Utilizes hardware intrinsics (AVX2, AVX-512) for lightning-fast brute-force vector comparisons (Cosine Similarity, Dot Product, Euclidean Distance).
• ⚡ Bare-Metal GPU Batch Search: Direct driver P/Invoke (nvcuda.dll and amdhip64.dll) offloading 2D grid batch scans (vector_batch_dot_scan_fp32) to NVIDIA RTX 4060 dGPU and AMD APUs without CUDA/ROCm SDK dependencies.
• 💾 Persistent Device VRAM Caching: Keeps database vectors resident in GPU memory across queries, bypassing host-to-device PCIe transfer penalties.
• 📈 Over 622 Million Vectors/sec: Delivers 10.1x sustained speedup over multi-core AVX-512 CPU execution for large batch queries.
• 🧠 LLM Optimized: Specifically designed to handle standard embedding dimensions (e.g., 1536 for OpenAI text-embedding-3-small or 768 for nomic-embed).
• 📥 Zero-Copy Memory Model: Efficiently manages large vector datasets in-memory using Span<T> and Memory<T>, minimizing allocations and GC pressure.
• 🤖 MCP Server Support: Built-in support for the Model Context Protocol, allowing seamless integration with AI agents and tools.
• 🛠️ Extensible Storage: Pluggable storage backends, starting with high-performance InMemoryVectorStorage.

📊 Performance Benchmarks

Benchmarked on .NET 10.0: AMD Ryzen AI 9 HX 370 (Zen 5 AVX-512) vs. NVIDIA GeForce RTX 4060 Laptop GPU (Ada Lovelace sm_89)

Installation

Glacier.Vector is available as a NuGet package. Install it using the .NET CLI:

dotnet add package Glacier.Vector

Quick Start

1. Initialize Index and Storage

using Glacier.Vector.Index;
using Glacier.Vector.Storage;

// Initialize storage for 1536-dimensional vectors
using var storage = new InMemoryVectorStorage(dimensions: 1536);
using var index = new VectorIndex(storage);

2. Add Vectors

float[] vector = GetEmbeddings("Hello, world!");
index.Add(vector, id: "doc_1", metadata: "Greeting message");

3. Search

float[] query = GetEmbeddings("Hi there");
var results = index.Search(query, topK: 5);

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

4. Bare-Metal GPU Batch Vector Scan

using Glacier.Vector.Compute;
using Glacier.Vector.Core;

// High-throughput 2D grid GPU batch search: computes Database * Queries^T
float[] db = LoadDatabaseVectors(count: 50_000, dim: 128);
float[] queries = LoadBatchQueries(batchSize: 32, dim: 128);
float[] scores = new float[50_000 * 32];

// Executes on NVIDIA RTX 4060 dGPU or AMD APU with persistent VRAM caching
GpuVectorAccelerator.BatchScanGpu(
    db, queries, scores, 
    vectorCount: 50_000, dimensions: 128, batchSize: 32, 
    target: GpuTarget.Auto
);

MCP Server Setup

Glacier.Vector includes a built-in MCP (Model Context Protocol) server host, allowing you to use your vector database as a tool for AI agents (like Claude or Antigravity).

1. Build the Server

First, build the host application in Release mode:

dotnet build src/Glacier.Vector.Host/Glacier.Vector.Host.csproj -c Release

2. Configure Your Client

Add the following entry to your mcp_config.json (usually located in %AppData%\Roaming\Claude\claude_desktop_config.json or your agent's config directory):

{
  "mcpServers": {
    "glacier-vector": {
      "command": "dotnet",
      "args": [
        "ABS_PATH_TO_REPO/src/Glacier.Vector.Host/bin/Release/net10.0/Glacier.Vector.Host.dll"
      ]
    }
  }
}

3. Available Tools

• add_vector: Adds a 1536-dimensional vector and associated metadata to the index.
• search_vectors: Performs a semantic search for the closest matches to a query vector.
• ping: Simple health check to verify the server is responding.

The server defaults to 1536 dimensions, optimized for modern LLM embedding models.

Performance

Glacier.Vector is designed to saturate your CPU's memory bandwidth during search operations. On modern hardware, it can scan millions of vectors per second using SIMD-optimized kernels.

Architecture

1. Compute Kernels: Static, SIMD-optimized methods for distance calculations.
2. Vector Storage: Manages the raw memory layouts of vectors and metadata.
3. Vector Index: Coordinates search operations and manages IDs/metadata mappings.
4. MCP Integration: Exposes the vector search capabilities via a standard Model Context Protocol interface.

Contributing

We welcome contributions! Please see CONTRIBUTING.md for details.

Credits

Developed by Ian Cowley and Antigravity (Google DeepMind).

License

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