RuVector is a high-performance, self-learning vector and graph database built in Rust, purpose-designed for AI and agentic systems. It combines HNSW vector search, graph neural networks, and local LLM inference into a single unified engine, delivering sub-microsecond latency and 10-100x performance improvements over Python and TypeScript alternatives. The project, released under the MIT license, has gained traction for its innovative 'cognitive container' concept -- single .rvf files that bundle vectors, models, and configuration into executable microservices that boot in 125 milliseconds. ## Vector Search and Graph Intelligence RuVector uses HNSW (Hierarchical Navigable Small World) indexing with SIMD optimization for vector similarity search at sub-microsecond latency. It also supports Cypher-language graph queries, enabling relationship traversal alongside vector retrieval. This dual capability means developers can build RAG pipelines that reason over both semantic similarity and structured knowledge graphs in a single query. ## Self-Learning Memory Unlike traditional vector databases that treat stored embeddings as static, RuVector includes GNN (Graph Neural Network) layers that learn from query patterns over time. The database gets smarter the more it is used, automatically optimizing retrieval paths and relevance scoring based on historical usage. It supports 46 attention mechanisms including flash attention and min-cut gating. ## Cognitive Containers (RVF) The RVF container format is one of RuVector's most distinctive features. A single .rvf file stores vectors, ships models, boots as a Linux microservice in 125ms, accelerates queries via eBPF, branches like Git at cluster granularity, and proves every operation through a cryptographic witness chain. This makes deployment as simple as distributing a single file. ## Local LLM Integration Through the ruvllm module, RuVector can run language models locally using GGUF format, eliminating the need for external API calls. This is particularly valuable for privacy-sensitive applications and edge deployments where network connectivity cannot be guaranteed. ## Distributed Architecture RuVector supports Raft consensus, multi-master replication, and auto-sharding for horizontal scaling. The engine also includes a PostgreSQL extension with 77+ SQL functions and pgvector compatibility, allowing teams to add vector capabilities to existing Postgres deployments. ## Compression and Efficiency Adaptive tiering provides 2-32x memory reduction, making it feasible to handle large-scale vector collections on commodity hardware. The sublinear solver algorithms operate at O(log n) complexity for sparse linear systems.