Try it with your own rules
Start from the preset below, adapt the questions, and inspect the typed answers. Run it live in the playground — it is the same request shape your application will send.
What Jev returns
Illustrative output for the preset below — run it live to get real values for your input.
Compare the attached 40-page supply agreement against our standard terms and list every clause that deviates on liability caps or termination rights.
Which model tier should handle this request?
Does this request need deep multi-step reasoning?
P(yes) = 0.97 — long document comparison across two clause types
How demanding is this task?
From one example to a reusable workflow
Name the routes as labels
Define each option by what it can handle, not by model name — "small: routine asks under 2k tokens" survives a model catalog change.
Gate on uncertainty
When the choice confidence is low, route to the safer tier by default — a router that knows it is unsure beats a confident wrong route.
Log route and outcome
Record the chosen tier, confidence and downstream quality. Re-tune when traffic or your model menu shifts.
Keep the criteria separate
| Check | What it measures | How to use it |
|---|---|---|
| Route · Choice | Which tier or handler should take this? | Keep option definitions stable; the probabilities map doubles as a fallback signal. |
| Reasoning · Yes/No | Does it need multi-step depth? | A noul gate catches "looks simple, isn't" requests before they hit a cheap model. |
| Complexity · Score | How demanding is it overall? | A score gives you a tunable knob between binary easy/hard — route mid-band to standard and watch quality. |
A single model for every request is easy to ship but expensive to operate
Rules catch obvious cases, while a general LLM is too slow and costly for a small routing decision. A decision model sits between them: typed answers, sub-second latency, pennies per thousand calls.
The route is a decision, not a vibe
Ask for the tier as a choice over named options, reasoning need as a probability, and complexity as a score — then branch in code. Your router becomes inspectable instead of a prompt that returns prose.
The production loop
Log the Jev route, confidence, final answer quality and any fallback to a bigger model. The routing policy improves when measured against what each tier actually delivered — revisit the questions as traffic and models change.
FAQ
How fast is a routing decision?
A single request with a few questions lands well under a second — small enough to sit before a generation call without dominating latency.
Can the router choose tools, not just models?
Yes — a choice question over handler names works the same for tools, queues or pipelines.
What happens when Jev is unsure?
Read the probability distribution: low top-probability means ambiguous input. Route to a safer tier or a human instead of trusting a coin-flip route.