Jev vs Classifiers, Embeddings and Zero-Shot Models

Last checked Sep 19, 2026 · Independent guide, not affiliated with TypeSafe AI

ANSWER

Jev is a general-purpose classifier you configure with words instead of training data: you describe the options and it returns calibrated probabilities, with nothing to train or host. A fine-tuned classifier can be cheaper and more accurate for one fixed, high-volume task where you have labeled data. Embeddings and zero-shot models sit in between.

Many Hacker News commenters’ first reaction to Jev was: “Isn’t this a classifier?” Mostly, yes. The interesting question is how it compares with the classifiers developers already use.

Four ways to classify text

Aspect	Jev	Fine-tuned classifier	Embeddings + small model	Zero-shot classifier (e.g. GLiClass)
How you define the task	Plain-language question and option descriptions	Labeled training data	Labeled data to fit a model on vectors	Label names and descriptions
Labeled data needed	None to start; some to pick thresholds	Hundreds to thousands	Dozens to hundreds	None
Change the labels	Edit the request	Relabel and retrain	Refit	Edit the labels
Hosting	TypeSafe’s API	You host it	You host or call an embeddings API	You host it
Probabilities	Calibrated by design	Often poorly calibrated without extra work	Depends on the model you fit	Scores, not necessarily calibrated
Understands long instructions and rubrics	Yes	No	No	Limited
Cost per call	$0.042 per million input tokens	Near zero at scale, after training	Low	Your compute
Latency	A few hundred ms over the network	Milliseconds locally	Milliseconds plus embedding call	Tens of milliseconds locally

Where Jev wins

No training loop. You can go from idea to working classifier in an afternoon, and change the categories tomorrow by editing text.
Rubrics and context. Jev reads instructions, criteria and structured state, so a question can depend on a policy document or an account record. A classic classifier only sees the input text.
Many small questions at once. One request can ask 20 different yes/no questions about the same input for about the price of one.
Calibrated uncertainty. Probabilities you can threshold for human review without a calibration step of your own.

Where a trained model wins

One fixed task at very high volume. If you classify millions of items a day into categories that never change, a small model you host can be cheaper and faster than any API.
Plenty of labeled data. With thousands of good labels, a fine-tuned model often reaches higher accuracy on its narrow task than any general model.
Offline or on-premises requirements. Jev is only available as a hosted service.
Latency budgets under the network round trip. A local model answers in milliseconds.

Embeddings: a middle path

Embedding the text and fitting a logistic regression or nearest-neighbor model on labeled examples is cheap and fast. It works well when categories correspond to topics. It struggles when the label depends on a subtle condition (“is the customer threatening to leave?”) or on information outside the text. Jev handles those better because the question itself is part of the input.

Zero-shot classifiers

Zero-shot models such as Knowledgator’s GLiClass score text against labels you supply without training. They predate Jev (GLiClass since 2024) and run on your own hardware, so they are the natural comparison for “classification without training data”. Jev’s differences are its ability to follow longer instructions and rubrics, its calibrated probabilities, and the three question types. Someone reposted GLiClass on Hacker News during launch week as an “open-source Jev”; they are related ideas, not the same model.

Combining them

The approaches are not exclusive. TypeSafe’s own autoresearch cookbook uses Jev to turn free text into numeric features and then trains a classical CatBoost model on top, keeping language understanding in Jev and task-specific fitting in a model you own. Another common pattern is to label data with Jev, review a sample by hand, then train a small, cheap model for the highest-volume path.

A quick decision guide

Your situation	Start with
New task, no labels, categories may change	Jev
Decision depends on rules, policies or account data	Jev
Fixed task, lots of labels, huge volume	Fine-tuned classifier
Topic-like categories, some labels, need low cost	Embeddings + small model
Must run locally with no training	Zero-shot classifier