Power State, Not Kernel
Same Query, Different Speed
4.6
×
Light-path kernel time across five RTX 4090 hosts
Heavy path stayed within 1–7%
Built for Exact Retrieval
An embedded vector store for Rust, Python and TypeScript. One durable file, metadata filters applied before ranking, exact results on CPU by default, optional GPU. Every number below links to its source and its boundary.
Start with the quickstart
Read how it works
Install 0.1.0-beta.2
Power State, Not Kernel
New in beta.2
Tiled GPU Top-k
4.5
×
Faster device top-k selection, RTX 4050 laptop
Results unchanged, checked against FP64
Every Number Traceable
Hashed Evidence Archive
1.8
K
1,833 retained evidence files
SHA-256 inventory, checked in CI
Install
Rust
cargo add qenloPython
pip install qenlo==0.1.0b2TypeScript
npm install @a3ro.dev/qenlo@betaGo, Kotlin and Swift are previews. Quickstart · SDK guides · Releases
Current research · GPU power states
The efficient kernel runs slow.
Our compact-row GPU path varied up to 4.6× across tested fresh processes and hosts, while a heavier path was more stable. A load-raising experiment reduced the device-time tail, but its predicted end-to-end speedup failed. We publish both results and the protocols behind them.