A self-learning vector database integrating GNN-driven search optimization, local LLM inference, Cypher graph queries, and a PostgreSQL vector extension, deployable from WASM embeddings to Raft-distributed clusters.
RuVector is a highly modular vector database built in Rust (v2.2.2, MIT licensed). Its core differentiator is "self-learning" capability—through SONA (Self-Optimizing Neural Architecture) and GNN-driven search learning, indexes auto-optimize with every query, forming a closed loop: query → response → feedback → learning signal (<1ms) → GNN index update.
Vector Retrieval Engine#
- HNSW + SIMD acceleration: Compatible with OpenAI / Cohere / local ONNX embeddings
- Hybrid search (RRF): Sparse + dense vector Reciprocal Rank Fusion (claimed 20–49% retrieval improvement, unverified by third parties)
- Graph RAG: Knowledge graphs + Leiden community detection + multi-hop queries (claimed 30–60% improvement, unverified)
- DiskANN / Vamana: Billion-scale SSD ANN search (claimed <10ms latency)
- ColBERT multi-vector: Token-level late interaction (MaxSim / AvgSim / SumMax)
- Matryoshka embeddings: Adaptive dimension funnel search
- OPQ product quantization (claimed 10–30% error reduction)
- TurboQuant: 2–4 bit KV-cache quantization (claimed 6–8x memory savings, <0.5% quality loss)
- LSM compaction: For write-intensive vector workloads
Graph Database#
- Cypher engine: Neo4j-like syntax with hyperedge support
- Graph Transformers: 8 validated modules covering physics, biology, manifolds, temporal, economics, etc.
- Hyperbolic HNSW: Poincaré ball space hierarchical search
Self-Learning System#
- SONA: LoRA fine-tuning + EWC++ memory retention (claimed <1ms adaptation)
- GNN search learning: Trains GNN layer per query, indexes improve automatically with use
- Cross-domain transfer learning and overnight auto-training
- 50+ attention mechanisms (count pending per-crate verification): FlashAttention-3, MLA, Mamba SSM, graph attention, hyperbolic attention, mincut-gated, etc.
Local AI Inference#
- ruvllm: Loads GGUF models with Metal / CUDA / ANE / WebGPU backends
- Sparse inference engine: PowerInfer-style, activating only needed neurons
- RuvLTRA model: Pretrained GGUF (claimed routing and embedding <10ms)
- Tiny Dancer: FastGRNN-based lightweight agent routing
Distributed & Consistency#
Raft consensus + multi-master replication + auto-sharding; Delta consensus (CRDT + causal ordering).
Cognitive Containers (RVF Format)#
Single .rvf file packaging vectors + model + Linux kernel (claimed 125ms boot), 5.5 KB WASM runtime for in-browser queries, COW branching, cryptographic witness chains, post-quantum signatures (ML-DSA-65).
PostgreSQL Extension#
230+ SQL functions as a drop-in pgvector replacement, with in-SQL sub-linear solvers (PageRank, conjugate gradient, Laplacian solver).
Deployment#
Single .rvf file → server / browser(WASM) / mobile / IoT / bare metal, with backend acceleration across Metal / CUDA / ANE / WebGPU / eBPF / FPGA. ESP32-S3 embedded platforms supported.
Experimental Modules#
Genomics (variant detection, protein translation, HNSW k-mer search), quantum coherence (ruQu, actual quantum relevance unconfirmed), consciousness metrics (IIT Φ, theoretical rigor unconfirmed), FPGA transformers, robotics, neural trading strategies, etc.
Installation#
curl -fsSL https://raw.githubusercontent.com/ruvnet/ruvector/main/install.sh | bash
npx ruvector
cargo install ruvector-cli
npm install -g ruvector
Caveats#
- Performance benchmarks (retrieval improvements, latency, memory savings) are self-reported in README; no independent third-party benchmarks found
- Maturity and practical utility of quantum coherence and consciousness metrics modules remain unverified
- No peer-reviewed publications found
- Associated product Cognitum claims CES 2026 Innovation Award, but CES 2026 has not yet taken place—requires verification