Zero-shot · your own categories · no API key

Jev Classifier

A free zero-shot text classification tool: type the categories you care about, paste any text, and the jev classifier returns the label that fits — with a calibrated confidence score and a full breakdown for every category, in about a second.

4 / 100
bugpraisefeature requestquestion

Zero-shot: no training data needed. Descriptive names classify best — “urgent bug” beats “p0”.

Fast tier — answers in about a second.

0 / 32,000

Your classification will appear here — winning label, confidence, and the score for every category.

The page runs in your browser and this site stores nothing; the text you submit is sent over HTTPS to the free, keyless classifier.dev API to compute the result.

How it works

How the Jev Classifier works

Three steps from a list of categories to a labelled answer.

1

Define your categories

Type your own labels — bug, praise, feature request, urgent, billing, whatever your work needs — or start from a preset. The jev classifier is zero-shot, so there is no training data to upload and no model to fine-tune: two to a hundred plain-language names are the entire setup.

2

Paste the text to classify

Drop in a support ticket, a review, a headline, a log line, or up to 32,000 characters of anything written. Pick the fast tier for a sub-second answer or the smart tier, which re-asks borderline cases with a reasoning model before committing.

3

Read the result — scores included

The winning label lands front and center with its calibrated confidence, and a bar for every category shows the full score distribution behind the decision. One tap copies a shareable summary; the same labels always route the same text the same way.

Features

What sets this jev classifier apart

Your own taxonomy, calibrated confidence, and two tiers of accuracy — no key, no account, no training set.

Your taxonomy, not a fixed one

Most text classification tools hand you a frozen label set. The jev classifier flips it: you name the categories, in your own vocabulary, and the model maps meaning onto them. Descriptive labels classify best — "urgent bug" beats "p0" — and you can retune the whole taxonomy in seconds.

Calibrated confidence you can act on

Every result carries a confidence you can put a threshold on: route anything above 0.9 automatically, queue the 0.5–0.9 band for a human, and drop the rest. The per-label score bars show the runner-up too, so you see when a text was really a coin flip between two categories.

Fast and smart tiers, no key required

The fast tier answers in roughly a second. The smart tier re-asks any answer below 0.7 confidence with a reasoning model, lifting accuracy on exactly the items that need it. Free use needs no account or API key — fair-use limits are counted per IP, like the rest of the free browser tools on this site.

Who it's for

Who uses the Jev Classifier

Anyone with more text than time to read it.

Support and feedback triage

Paste a ticket into the jev classifier with labels like bug, praise, feature request, and question, and it lands in the right queue before anyone reads it. Fifty reviews become fifty sorted rows; the angry ones about checkout get counted, not lost in a scroll of five-star noise.

Content routing and moderation

User submissions, comments, and community posts need to reach different teams or pages. Classify them into your own buckets — support, sales, spam, off-topic — and let the confidence score decide what deserves human eyes. Retrain nothing when the buckets change; rename the labels and classify again.

Logs, streams, and inbox zero

Error buckets, inbound email subjects, changelog entries nobody reads line by line: classify first, read the survivors. The jev classifier gives deterministic routing for pipelines — the same input takes the same branch on every run — and the underlying classifier.dev API accepts up to 1,000 texts per call when you outgrow pasting one at a time.

Worked examples

Jev classifier examples

Real texts, real categories, and the answers this text classification tool returned:

TextCategoriesWinnerConfidence
The checkout button does nothing when I tap itbug · praise · feature requestbug100%
Absolutely love the new dark mode — great work!bug · praise · feature requestpraise100%
Request: please add CSV export to the reports pagebug · praise · feature requestfeature request100%
The package arrived three weeks late and support never repliedpositive · negative · neutralnegative100%
It compiles, I guessbug · praise · feature request · question · otherother82%

Real responses from the jev classifier (fast tier, jev-1.13.0, captured 2026-09-20). Confidence is the model’s calibrated certainty — note how ambiguity in "It compiles, I guess" pulls it down.

What Is Zero-Shot Text Classification?

Classic text classification is a training problem: collect a few hundred labeled examples, fine-tune a model, deploy it, and repeat whenever the labels change. Zero-shot classification skips all of it. The model already understands language; your label names are themselves descriptions of a category, so it scores how well the meaning of your text matches the meaning of each label and returns the ranking. The jev classifier is that idea wrapped in a form anyone can use — define the categories in plain words, paste text, get an answer.

The strength and the weakness are the same property. You can change your taxonomy on a whim — rename a label, add a bucket, split one category into three — and the next classification already respects it, which is ideal for support queues, moderation, and one-off research questions. What you give up is the last few points of accuracy a fine-tuned model earns from seeing your exact data. For most sorting work, calibrated confidence closes that gap: anything the model is unsure about surfaces with a low score instead of a confident wrong answer.

It pairs naturally with the other free browser tools on this site. Prep and clean inputs with the Meta Glasses Converter, classify the stream with the jev classifier, and when the work turns to protocol plumbing, the A2A Protocol Validator checks agent cards the same keyless way.

Calibrated Confidence — and When to Trust It

A classifier that is right 90% of the time is useless if it is wrong confidently. Calibrated confidence fixes that: the score the model reports matches how often it is actually right, so a 0.9 answer fails about one time in ten. That is what makes thresholds safe to automate. Route everything above 0.9, send the 0.5–0.9 band to a human, and treat anything lower as a coin flip the model never should have been trusted with.

The jev classifier shows the full per-label distribution, not just the winner, because the runner-up often carries the story: a ticket scored 0.55 bug and 0.40 feature request is a different animal from one scored 0.95 bug outright. The smart tier automates the pattern — any answer below 0.7 confidence is re-asked with a reasoning model, and results carry an escalated flag when that happened. On public benchmarks that re-asking moves borderline accuracy by several points, which is exactly where the interesting cases live.

Deterministic routing is the quiet superpower: the same text with the same labels takes the same branch on every run, unlike prompting a chat model and hoping it feels consistent today. For pipelines that must not drift, that matters more than a benchmark point or two. More engineering deep dives live on the blog, and the A2A documentation covers the protocol this site is built around.

FAQ

Jev classifier FAQ

Zero-shot basics, confidence, privacy, and label craft.

What is the jev classifier?

The jev classifier is a free web tool for zero-shot text classification. You define a list of categories, paste any text, and it returns the category that fits best along with a calibrated confidence score and per-label scores. Under the hood it calls the free classifier.dev API, which runs the Jev classification model; the page itself is just a browser interface, so there is nothing to install or sign up for.

What is zero-shot text classification?

Zero-shot classification means the model assigns labels it was never explicitly trained on, with no example dataset from you. Instead of learning "these 500 emails are spam", it compares the meaning of your text against the meaning of your label names — so a category list like billing, bug, praise is all the setup required. You trade some fine-tuned accuracy for the ability to change categories in seconds.

Do I need an API key or an account?

No. The tool works immediately in the browser with no key, no login, and no billing. Free fair-use limits are counted per IP address: 3,000 fast-tier classifications per minute (20,000 per day) and 200 smart-tier per minute. That is far beyond what clicking a web page can consume.

What does the confidence score mean?

Confidence is the model’s calibrated certainty that the winning label is correct, from 0 to 1. Because it is calibrated rather than raw, a 0.85 really is right about 85% of the time, which lets you act on thresholds: auto-route above 0.9, review the middle band, and escalate below it — or switch to the smart tier, which automatically re-asks answers under 0.7 with a reasoning model.

Is my text private?

The page runs in your browser and this site stores nothing. Classification is a model’s job, so the text you submit is sent over HTTPS to the classifier.dev API to compute the result — that is how the answer gets back. Treat it like any cloud service: fine for tickets, reviews, and notes; think twice before pasting trade secrets or personal data.

How many categories can I use, and how do I write good ones?

Between 2 and 100 labels, and descriptive beats terse: "urgent bug" classifies better than "p0", "feature request" better than "fr". Keep labels mutually exclusive when you want one clear winner, add an "other" bucket to catch stragglers, and reuse the same list across runs so routing stays consistent.