Capabilities
Everything Jev can do revolves around three question types. One request may carry many questions (the docs call this parallel questioning, including “speculative” ones), each answered independently.
| Type | Asks | Answer contains | Confidence |
|---|---|---|---|
| Choice | Pick one from ≤255 defined options | choice + probabilities per option |
✅ |
| Score | Rate against 2–10 ordered levels | score (probability-weighted mean, may be fractional) + legend + probabilities |
✅ |
| Noul | Yes/no question | noul (probability of “yes”, 0–1) |
❌ |
Fields common to all types:
instructions: the question itself, in natural languagecriteria: your enumerated definition of options/levels/criteria — this is where the “type safety” comes from: you define the answer space
Parallel questions
Section titled “Parallel questions”A request’s questions map can hold many questions sharing one state. The official cookbook example: 13 regulatory questions over the GDPR article in one call were 12.2× cheaper and 10.0× faster than one call per question, with unchanged answers (official figures, verified 2026-09-21 via the official llms.txt index).
The docs also support speculative fan-out: send the questions that might matter and let your code decide which answers to use — ideal for judgment trees with many branches.
How to read confidence
Section titled “How to read confidence”The docs are explicit: confidence is not a probability and not accuracy — it is a statistic of the answer distribution’s shape. The higher the concentration, the higher the confidence. The intended split:
- The answer tells you what (
choice/score/noul) - The confidence tells you whether to act — route low-confidence cases to humans or a fallback
There is an official pattern for exactly this, Confidence-gated routing (via the llms.txt index); all four of our use cases build on it.
Usage and cost
Section titled “Usage and cost”Every response includes usage: { input_tokens, output_tokens }. Billing (official models page, verified 2026-09-22):
- Input $0.042 per million tokens; output is free
- ≤ 64k tokens per request;
state+ the longest single question ≤ 32k tokens - Rate limits: 250k tokens/s + 1200 requests/min (429-triggered; officially noted as subject to change)
When not to use it
Section titled “When not to use it”See Limitations: text generation, open-domain QA, and long reasoning chains are all out of scope — including a quick checklist for judging fit.