Vector database that actually learns from your queries - stores embeddings, queries with Cypher like Neo4j, scales with Raft consensus, and improves search results using Graph Neural Networks

Most vector databases are glorified static indexes - you dump embeddings in, get similarity searches out, repeat forever. RuVector flips this: every query trains a Graph Neural Network that makes future searches smarter. It’s like having a Neo4j that learns patterns in your data and gets better at finding what you actually want, not just what’s mathematically closest.

Built in Rust with some serious engineering: Raft consensus for horizontal scaling, 40+ attention mechanisms, local LLM inference via ONNX, and even runs in browsers via WASM. The Cypher query support means you can ask complex relationship questions, while the GNN layers quietly learn which connections matter most for your specific use cases. Ships as a single-file microservice that boots in 125ms, or deploy it as a distributed cluster.

At 354 stars, this is still flying under the radar but the architecture is genuinely novel. If you’re building anything with embeddings beyond basic similarity search - RAG systems, recommendation engines, knowledge graphs - this could save you from rebuilding indexes every time your data relationships evolve.

⭐ Stars: 354
💻 Language: Rust
🔗 Repository: ruvnet/ruvector