What is jevish and what does it do?

A zero-dependency JavaScript library for semantic pattern matching and zero-shot judgment that behaves like Jev in local CPU cache under 0.05 ms and can speculatively escalate tough cases to cloud Jev when you enable cascade.

How do I install and call it?

npm install jevish, then either pass semantic pattern handlers such as a bug rule with a >0.8 threshold or an array of labels for zero-shot classification. The winning label is the return value, and that is described as the whole API.

What does the cascade mode do?

With cascade: true unambiguous queries resolve in CPU cache at about 0.05 ms and $0 cost while tough edge cases are speculatively escalated to TypeSafe's cloud Jev. The README reports a 76% fast-path rate on banking intents.

Is the in-tree engine a neural model?

No. The README calls it deliberately primitive and dumb by design: tokenization, stemming, subword n-grams and a compact lexicon of semantic synonym clusters, with no 50MB weights, no ONNX runtimes and zero network calls.

How accurate is jevish in practice?

The README's benchmarks, N=100 per task, report 86.0% on banking77 intent routing in-tree and 99.0% with cascade, and 72.0% versus 98.0% on sms_spam. These are the project's own runs against Jev API baselines.

How do I teach it domain jargon?

Pass custom clusters, for example shipping mapped to tracking, parcel, courier, fedex, ups and package. The README frames this as zero-retraining extensibility: no fine-tuning and no vector database required.

What are the limits of the in-tree engine?

It only commits when the winning score and margin are high (score >= 0.70, margin >= 2.0). Genuinely ambiguous or out-of-vocabulary inputs need the cascade, and the digest truncates the benchmark table mid-suite.

Jev directory

CommunityJev Repos & SDKs

Jevish

Zero-dependency JavaScript library for semantic pattern matching and zero-shot judgment: the in-tree engine behaves like Jev in local CPU cache under 0.05 ms and speculatively escalates ambiguous inputs to cloud Jev when cascade is enabled.

See highlights Open github.com

Stars: 2
Language: JavaScript
License: MIT
Reading time: 2 min
Published by: Community
Author: hemanth
Last verified: Oct 2026

hemanth/jevish

User repository on GitHub

View on GitHub

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity catalog-import logicrw javascript classification

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: JavaScript
Platforms: Cloudflare Browser
Capabilities: Open sourceRuns locally

Highlights

One 8.6 kB zero-dependency engine: tokenization, stemming, subword n-grams and semantic synonym clusters, with no tensor weights or ONNX runtime.
The in-tree engine commits only with score >= 0.70 and margin >= 2.0 over the runner-up; otherwise cascade: true escalates to cloud Jev.
jevish.detailed() exposes calibrated probabilities: for a database connection pool error it reports score 0.96 with bug 0.96 and feature 0.04.
README benchmarks (N=100): banking77 intent routing 86.0% in-tree at 0.05 ms and 99.0% cascaded, versus 100% for the Jev API at 139 ms.
Auto-currying API: jevish(['bug','feature','billing']) returns a reusable classifier you can map over a list of tickets.

Quickstart

import jevish from 'jevish';

const category = await jevish('Can you provide an invoice for last month?', [
  'bug',
  'feature',
  'billing',
]);

Watch out

Grounded on the truncated README and GitHub metadata reviewed 2026-10-15; the benchmark table is cut off after the spam row and the numbers are the project's own runs with no independent verification. The README calls the in-tree engine intentionally primitive and dumb by design.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

The Open Reimplementations: Local Models That Speak System One

Sep 22, 2026

Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.

Jev Repos & SDKs#open-models#research#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai