Most vector stores return the same nearest neighbors every time; this project treats the index itself as a learning substrate. RuVector layers a GNN and lightweight runtime adaptation on top of HNSW so search results improve from usage patterns, session context, and micro-LoRA updates — all on your hardware.
What Sets It Apart
- GNN-on-index (self-learning reranking): the engine re-ranks HNSW candidates with a small GNN that updates from query/feedback traces, so relevance improves over time without full retraining — useful when user behavior and context matter.
- Local AI runtime + quantization: ships ruvllm and tooling to run GGUF models (Metal/CUDA/WebGPU/WASM), plus aggressive KV/adapter quantization so you can do RAG and inference locally without per-query cloud bills.
- Unified vector+graph substrate: Cypher-style graph queries, hyperedges and temporal graph modules let you express multi-hop reasoning and graph-RAG workflows without a separate graph DB.
- RVF cognitive containers & Postgres extension: deploy as a single .rvf artifact or drop into PostgreSQL (pgvector-compatible functions), easing integration into existing stacks and edge deployments.
Who it's for & tradeoffs
Great fit if you need an on-prem / edge-capable semantic store that adapts from user behavior (agent memory, personalized search, multi-agent orchestration) and you can run Rust-native or WASM components. Also valuable when you want tight Postgres integration or local LLM inference without cloud costs.
Look elsewhere if you need a lightweight, minimal-maintenance hosted vector service — RuVector is feature-rich and architecturally complex, with many experimental/advanced modules (RVF containers, formal proof gates, custom hardware claims) that increase operational surface and require evaluation for production hardening and security reviews.
