Old-school classifiers are still generally better than Jev [P]

There’s a lot of hype right now around decision models like Jev. But in most real use cases, you have some labeled data, and can use it to personalize your model.

So we tested the old-school approach to classification and regression:

• Embed your inputs with an open-source or cheap embedding model
• Fit a simple sklearn model on top using your own labels

Turns out this approach was cheaper and almost always significantly more accurate (37% error reduction against Jev).

Write-up and Code

Write-up: https://blog.voyageai.com/2026/10/27/how-to-create-your-own-jev-using-voyage-models/ (Disclosure: I work at Voyage)

Code: https://github.com/mishelsikder/mongo-predict

Curious whether others have seen the same thing, or found cases where zero-shot performs better.

Edit: To be clear, the gains here come from having training labels, which only applies to some use cases. Jev doesn’t use training labels, so it's unsuprising that this approaches wins when you have them and loses when you don't.