arrowspace-rs
Core engine in Rust. Builds λτ-indexed signal graphs and serves dispersion-model analytics and search for your vectors.
★ 41
AI research engineering · London, UK
Graph-native vector search.
Spectral generative models.
We build ArrowSpace, a Rust engine that indexes embeddings as signal graphs and searches them through graph Laplacians and per-feature energy dispersion. Around it we maintain Python bindings, release-gated benchmarks, diffusion models for sound, and the SmartCore machine-learning library.
- Rust-first
- Science-driven software
- Minimal by design
Flagship
ArrowSpace — vectors as graphs
ArrowSpace treats embeddings as signals on a graph. Search walks a graph Laplacian guided by λED, an energy-dispersion network over features. Results then respect dataset structure, not only distance. Graph analysis, vector search, and energy-distribution stats ship in one package.
pyarrowspace
Python bindings for ArrowSpace. Build signal graphs, run spectral queries, and analyse datasets from NumPy workflows.
★ 16
arrowspace-benches
Benchmark suite for the engine. Runs at every release to track recall and latency against baselines.
Generative audio
Latent Sound Diffusion
LSD is a latent diffusion model for sound production. A frozen ArrowSpace prior (LF, λED) defines the semantic manifold. An EnCodec encoder produces one-dimensional audio latents. Diffusion runs on those latents while a trainable graph decoder reconstructs waveforms on the feature-space manifold.
The model is a tool for producers, not a research claim. Train fast on your own samples. Every run differs by design: seeds stay fresh and latent noise injection is a taste knob.
Spectral Composition
A new paradigm for sample-based music creation via latent diffusion
DOI 10.5281/zenodo.21950475Composition as navigation of a compressed spectral manifold, estimated once from the producer’s whole sample library with ArrowSpace.
Stewardship
SmartCore — machine learning in Rust
We maintain and extend SmartCore, a comprehensive library for machine learning and numerical computing in Rust. Regression, classification, clustering, and matrix computation, built from first principles.
More from the lab
Research code
Smaller tools and paper implementations. Most of them Rust, all of them minimal.
nanogpt-rs
Karpathy’s nanoGPT ported to Rust with the Burn framework.
RUSTtauformer
Transformer built on taumode attention. Topological search attacks high-dimensionality. Companion paper included.
RUSTkalman_clustering
Cluster any vector space with a Kalman filter.
RUSTtopological-pagerank
MRR-Top0: ranking over the corpus graph that scores the whole top-k list, not just the first hit.
TEXoptical-embeddings-rs
Optical embeddings in the DeepSeek-OCR style.
RUSTconstraint-decoding-trie-static
STATIC constrained decoding: a prefix trie flattened to a CSR matrix, enforced by a vectorised kernel.
RUSTsolid-community-rs
Rust port of the Solid Community Server: LDP, WAC, and WebID-TLS specifications.
RUSTjavelin-tui
Terminal tool for large embedding sets on Lance files, with graph-based search.
RUSTgenegraph-storage
A small storage layer for large matrices and embeddings.
RUSTdeep-delta-rs
Deep Delta Learning, arXiv:2601.00417, implemented in Rust.
RUSTvibrational-deduction-transformer
Deductive transformer guided by Laplacian eigenbases, treated as harmonics of a vibrating system.
PYTHONSupport
Sponsor the work
Maintenance of ArrowSpace, SmartCore, and the research stack runs on personal time by default. Sponsorship tiers fund open-source upkeep, release benchmarks, and paper reproduction work. Sponsors also receive the Science-Driven Software newsletter.
If your organisation depends on vector search or LLM tooling, consider a tier.