FerresDB is a persistent vector search engine for semantic search and RAG. Built in Rust with HNSW, it features a built-in Dashboard, gRPC/REST APIs, WebSockets, and WAL for crash recovery. Fast, reliable, and observability-ready.

I developed FerresDB because I wanted a vector engine that was blazing fast (thanks to Rust!) without sacrificing the developer experience. It comes out of the box with a full-featured Dashboard and gRPC support. I’m evolving the project step by step and would love to get your feedback on its performance and interface. It’s already available for testing via Docker!

I’ve just released a series of fundamental improvements to FerresDB, focused on low-level performance and native integration with AI ecosystems.

What’s new:

🔌 Embedded MCP (Model Context Protocol): Native support via STDIO. It’s now possible to connect the database directly to Claude Desktop or Cursor IDE.

⚡ SIMD-Accelerated Kernels: Implementation of distance kernels (Euclidean/Dot Product) in Rust using AVX2 and SSE4.1 instructions, with runtime detection.

🔍 Native HNSW Pre-filtering: Metadata filtering integrated directly into graph traversal, ensuring precision and returning the exact requested limit.

🏢 Logical Namespaces: Native multitenancy support, allowing data from multiple clients to be isolated within the same physical collection efficiently.

📊 Real-time Analytics: Updated dashboard with time-series charts for P95 latency and ingestion throughput, plus a hardware acceleration indicator.

📦 Storage Optimization: Added Zstd compression for the WAL and support for binary snapshots via bincode for ultra-fast loading.

🔄 Auto-Reindex & TTL: New background worker for automatic index compaction and support for Time-to-Live data expiration.

The project continues to evolve as a lightweight and resilient solution for vector search infrastructure.