Decide
A System One model in the browser.
Fast, typed decisions from text, on the user's device, with confidence you can trust. When it isn't sure, it asks an LLM and learns from the answer.
the app crashes when I open settings
Sure: answered on this device
A few lines of code
import { createDecider } from "decidekit";
const d = await createDecider();
const { choice, confidence } = await d.decide(
"my card still hasn't arrived",
["card_arrival", "lost_card", "refund"],
);
// choice: "card_arrival", typed as your labelsWatch it learn from an LLM
It starts out knowing only twelve label names. What it's sure of, it answers itself; the rest goes to an LLM, and every eight answers it retrains. The LLM's answers were recorded once, so the demo is free to watch.
What's a System One model?
Kahneman called fast, intuitive thinking System 1, and slow, careful thinking System 2. A System One model is the fast kind: it doesn't write anything, it reads the text once and picks one of the options you give it. That makes it quick, cheap and safe to wire into code.
Think of Decide as a local Jev. TypeSafe's Jev is a System One model behind an API; Decide runs on the user's device, so each decision is free, private and a few milliseconds. Jev hands back a confidence and leaves the unsure cases to your code. Decide can send them to an LLM, its System 2, and learn from the answer, so it needs the LLM less over time.
Further reading: System One models (TypeSafe) · Introducing Jev · Confidence in Jev · Thinking, Fast and Slow
Typed choices
The answer is always one of your labels, typed as their literal union. No parsing, no made-up labels.
Confidence you can trust
Trained heads are calibrated: an answer at 0.9 is right about 90% of the time, so a threshold means what it says.
Learns from an LLM
Wrap your LLM call in withFallback: unsure texts go to it, its answers become training data, and it gets asked less.
Runs in the tab
A 34 MB model in a web worker, cached after the first visit. It adds 3.6 kB to your page bundle.
Trains on the device
Train a head from a few examples per label in seconds, without sending the text anywhere.
Browser, Node, React, switch
The same API everywhere, React hooks, and a decider for the switch harness.
What it's for
Cut the LLM bill
Answer the easy majority on device; send only unsure texts to an LLM, which teaches it as it goes.
Route support messages
Billing, bug, account: route a message, and show the right help while the user is still typing.
Command palettes
"Make it darker" to theme.dark, in milliseconds and offline.
Games
Read typed player chat and commands with no server.
Sensitive text
Health, finance and internal tools: training and deciding stay on the device.
Agent harnesses
A free decision model for switch, with an LLM as its teacher.