Vector Search

Two vector search systems in one skill, and they're not interchangeable. The embeddings_* tools handle large corpora (millions of vectors) with HNSW indexing that's 150x to 12,500x faster than brute force, depending on size. The ruvllm_hnsw_* tools are a WASM router capped at 11 patterns for sub-millisecond routing, not corpus search. If you're memory constrained with 5,000+ vectors, RaBitQ gives you 32x memory reduction via 1-bit quantization. The hyperbolic option maps hierarchical data to Poincare ball space instead of using cosine distance. Know which path you need before you start, the capacity limits are real and the use cases don't overlap.

skills.sh installs: 752
GitHub stars: 81.4K
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add ruvnet/ruflo --skill vector-search --agent claude-code

Installs into .claude/skills of the current project.

Files

Vector Search

Two distinct vector-search paths live in this plugin. Pick the right one — they're not interchangeable.

Path	Tool family	Backing	Capacity	Latency
Large-scale corpus	`embeddings_*`	`@claude-flow/memory` HNSW (Rust/Native)	large corpora (measured here only at N=5k and N=20k)	~1.9× at N=20k, ~3.2×–4.7× at N=5k vs brute-force (measured; recall@10 ≈ 0.99). ANN wins above the crossover
Hot-path router	`ruvllm_hnsw_*`	WASM-backed router (v2.0.1)	~11 patterns max (`ruvllm-tools.ts:58`)	sub-ms; designed for high-priority routing, not corpus search

The speedups in the Performance section apply to the large-scale embeddings_search path (the older "12,500×" headline was a brute-force-fallback artifact and is retired). The WASM router is not that path.