ArrowSpace
An open-source library for spectral vector search: it indexes embeddings with bounded spectral scores derived from a graph Laplacian, and answers queries by blending geometric similarity with corpus structure.
ArrowSpace implements spectral indexing for vector similarity search: it builds a graph over embeddings, computes bounded spectral scores (λτ in [0, 1)) from Laplacian smoothness and dispersion statistics, and retrieves with a runtime blend of cosine similarity and spectral signal. See spectral vector search for the method.
Installation
Python (pyarrowspace):
Rust (arrowspace-rs):
Versions, release notes, and platform support are published on the repositories; consult them before pinning a version.
Capabilities
- Graph analysis and spectral indexing over embedding datasets, with parameterised graph wiring.
- Bounded spectral scores (λτ in [0, 1)) per item, comparable across collections and model updates.
- Tau-modulation at query time to blend spectral and cosine-like behaviour.
- Energy-distribution statistics that support re-ranking, thresholding, and monitoring workflows.
Evidence and evaluation
The method is published academically in the ArrowSpace paper and the Energy Dispersion Networks paper. Retrieval quality is evaluated against cosine baselines with a published protocol in Engineering 001 (CVE corpus and TREC-COVID). A live industrial deployment is the CVE Search Engine demo.
Repository
| Distribution | Language | Install | Repository |
|---|---|---|---|
| pyarrowspace | Python | pip install arrowspace |
tuned-org-uk/pyarrowspace |
| arrowspace-rs | Rust | cargo add arrowspace |
tuned-org-uk/arrowspace-rs |
Positioning
ArrowSpace complements standard similarity search rather than replacing it. It adds build and storage cost, and its benefit depends on graph construction, data distribution, and query type; see spectral search vs HNSW for the trade-off discussion.