← Back to the catalog

Bekko

Local Bekko System One decision provider: fast select, boolean and ordinal decisions from a small on-device model

Details:

DecisionsFirst-party

npmRepositoryHomepageReport an issue

Install

alisio install npm:@alisio/plugin-bekko@0.1.0

Metadata

Package	@alisio/plugin-bekko
Version	0.1.0
Published	Nov 11, 2026
Author	Alisio
License	MIT
Categories	Decisions
Size	81.7 kB
Dependencies	3
Peer dependencies	1
Downloads (last month)	—
Source	First-party
Keywords	alisio-plugin, bekko, decisions, system-one, onnx, local

Security note: Plugins run in-process with your user privileges and are not sandboxed. Review the source before installing a third-party plugin.

Español. Both READMEs must be updated together.

A local "System One" decision provider for Alisio. Features that ask Alisio for a select, boolean or ordinal decision get a fast answer from Bekko, a small typed-decision model that runs on your CPU: no LLM call, no Python, no GPU, and no data leaves your machine once the model is downloaded.

This is an independent plugin, not affiliated with the Bekko project. Bekko is optional: Alisio works without it, and every feature keeps its own fallback when no provider answers.

Requirements

• Alisio with the decisions API (>=0.9.0 <0.12.0); on older hosts only the commands load.
• Disk: about 200 MB for the default 68m model (about 33 MB for 17m), plus the npm dependencies (ONNX Runtime and Transformers.js, roughly 450 MB installed).
• Memory: the Alisio process grew to about 1.6 GB while the 68m model was loaded (measured on Linux x64); plan for that much free memory.
• Network only for the first download, from Hugging Face.

Install and activate

alisio install npm:@alisio/plugin-bekko

Then, inside Alisio:

1. /bekko:activate asks Alisio to make Bekko the active decision provider. Alisio asks you to confirm and saves decisions.provider itself; it never replaces a different provider you configured. On a host without that capability, set decisions.provider = "bekko" by hand.
2. Once active, the model downloads and loads in the background. Until it is ready, Alisio keeps using its normal fallback. /bekko:status shows progress.
3. Optional: /bekko:setup downloads and loads the model now, without changing the active provider.

Installing the package alone downloads nothing and starts nothing.

Commands

Command	What it does
/bekko:status	Model state (inactive, loading, ready, failed) with its detail, and the active decision provider.
/bekko:setup	Downloads (first run) and loads the model now; reports the download size.
/bekko:activate	Asks Alisio to make Bekko the active decision provider (Alisio asks you to confirm).

How it works

• Each request (DecisionRequest v1: up to 16 decisions, 32 KiB) is mapped to Bekko's typed questions: select to a choice among 2 to 20 options, boolean to a yes/no question and ordinal to a score over 2 to 10 levels.
• The model runs with ONNX Runtime Web on its WebAssembly backend (CPU) through jev-web. Loading the 68m model took about 7 seconds the first time and 2.5 seconds from the local copy on a desktop CPU; a three-decision request answered in about a tenth of a second.
• The provider never waits for the model: while it loads, a request fails fast as "not ready" and Alisio uses its fallback. Answers whose probabilities are invalid, or whose input the model had to truncate, are rejected instead of being guessed.
• confidence is the model's own probability for its answer. It is not calibrated, so keep Alisio's decisions.minConfidence threshold.

Model download and integrity

• Files come only from fixed revisions of hotchpotch/bekko-system-one-v0-68m (or -v0-17m) on Hugging Face: manifest.json, tokenizer.json and model.onnx.
• Every file is checked against a size and SHA-256 digest built into this plugin version before it is used, and is written to the cache only after it passes. Redirects are followed only within Hugging Face's domains (huggingface.co, hf.co).
• The cache lives in Alisio's per-plugin cache directory (or ALISIO_BEKKO_CACHE_DIR, or ~/.alisio/cache/bekko). A damaged cached file is detected and downloaded again. Delete the directory to free the space.

Limitations

• Model license. The Bekko model card says the license of the released weights "remains to be finalized" and does not assign one. This plugin does not ship the weights; it downloads them from Hugging Face when you activate it. Review the model card before using Bekko in a product.
• Bekko v0 is trained mainly on English. Decisions on other languages work but are less reliable.
• Inference runs on the main thread of the Alisio process. A request that exceeds its deadline is reported as a timeout, but the ONNX computation already started still finishes in the background.
• The 400M variant is not offered (1.4 GB download).

Advanced use

import { createBekkoPlugin } from "@alisio/plugin-bekko";

export default createBekkoPlugin({ variant: "17m" }); // smaller, lower quality

createBekkoPlugin also accepts cacheDir and an engineFactory (for tests or a custom runtime).

Development

pnpm --filter @alisio/plugin-bekko test
pnpm --filter @alisio/plugin-bekko build

Unit tests use a fake engine and a fake network; they never download the model.

Sources

• Bekko System One: https://github.com/hotchpotch/bekko-system-one (runtime code MIT) and https://huggingface.co/hotchpotch/bekko-system-one-v0-68m.
• jev-web (MIT), ONNX Runtime Web (MIT), Transformers.js (Apache-2.0).

License

MIT for this plugin. The Bekko model weights are not part of this package and carry their own terms (see Limitations).