A vector database that claims to learn while you query it

What it does RuVector is a Rust-based vector database with a sprawling feature list: HNSW indexing, Cypher graph queries, local GGUF inference, PostgreSQL extension, WASM builds, and a “SONA” self-optimization engine that feeds query feedback into GNN layers. It packages as a single .rvf file or runs via npx ruvector.

The interesting bit The project treats vector search as a continual learning problem rather than a static indexing one — every query becomes a training signal for LoRA fine-tuning. Whether that loop actually closes in production, or just closes in the README, is the open question.

Key highlights

• Ships as PostgreSQL extension with 230+ SQL functions, or 58 KB WASM for browsers
• Local LLM inference via GGUF with TurboQuant KV-cache compression (2–4 bit)
• Graph RAG with Leiden community detection and hyperedge support
• Raft consensus, multi-master replication, and Git-like COW branching for vectors
• Claims <1 ms adaptation latency and <10 ms DiskANN query latency on SSD

Caveats

• README is heavy on comparison tables and light on reproducible benchmarks
• “CES 2026 Innovation Award” badge links to the project’s own marketing site
• Many features (quantum coherence, genomics, “burst scaling”) are listed but not demonstrated
• The repo appears to be primarily authored by a single maintainer with heavy self-promotion

Verdict Worth watching if you’re building edge-first AI infrastructure and can tolerate alpha-grade ambition. Skip it if you need battle-tested vector search today — pgvector and Qdrill have fewer asterisks.