Route prompts to the right model with Jev

Can Jev choose which model handles a prompt?

Yes. TypeSafe lists model routing among its example use cases: a custom router that chooses which language model receives each prompt. Its intent-routing pattern shows the shape. One Choice question classifies the incoming request, and ordinary code sends each class to a handler: a database lookup, a specialist model or a human agent.

Jev does not write the reply. It makes the small decision in front of the expensive one, so that a routine prompt does not occupy a reasoning model and a hard one does not reach a model that cannot solve it.

Open the editable routing template. It contains one fictional engineering prompt and four handlers. Loading it creates a draft; a check starts only when you request it.

Describe handlers by what they are for

Model names change; responsibilities last longer. The template describes each destination by the work it suits, and your code maps the label to whatever model currently fills that role.

Change one thing and compare

1. Run the template and note the selected handler and the other probabilities.
2. Replace the prompt with “What are your support hours?” and run it again.
3. Try “Cancel my contract and refund this year’s fees.” Decide beforehand whether your policy lets software act on it.

No expected answer is prefilled, and a few runs do not measure routing quality. To evaluate a router, label a few hundred real prompts with the handler you would have chosen and compare, including what each mistake costs.

Use confidence as the fallback

In TypeSafe’s intent-routing example, the code checks confidence before it looks at the label: below a floor, the request goes to a person whatever the selected class. The confidence-gated routing pattern adds that the floor should follow the stakes, low for a harmless action and much higher for one that moves money. Pick your numbers from your own labelled prompts; the documented values illustrate the pattern and are not recommendations. Probability versus confidence explains the two signals.

What about trimming context?

A related idea asks a yes/no question per earlier message, such as “is this still needed?”, and drops the rest. Be careful. One community author reported on 18 September 2026 that a naive relevance filter gave no gain at best on a long-memory benchmark and a large drop at worst. That is a single author-reported experiment with one filtering method, but it is a useful warning: a dropped message cannot be recovered downstream. Measure answer quality with and without the filter before relying on it.

Routing mistakes are quiet, because a weaker model still produces fluent text. Log every decision, review a sample, and see where Jev’s limits are.