choiceup to 255 options
One of a known set
billing0.72
shipping0.20
other0.08
Σ = 1.00
choice · probabilities · confidence
Always add an escape option like “other”.
Cheatsheet
Jev on one page.
Three question types
score2–10 levels
A place on a described scale
0
1
2
score = 1.28
score · probabilities · confidence
Describe situations, not degrees.
noulone condition each
A clean yes or no
P(yes)0.93
0 no0.5 equally likely1 yes
noul (0–1) · no confidence
0.5 means “equally likely”, not “medium”.
Limits and price
- per 1M input tokens
- $0.042per 1M input tokensoutput is free
- tokens per request
- 64ktokens per request32k for state + longest question
- requests / min
- 1,200requests / min250k tokens / sec
- only
- Textonlystring, object or array
jev-latest and jev-preview both point at jev-1.13.0. Pin the version once you tune thresholds.
Patterns
12.2× cheaper
Speculative fan-out
Ask everything in one request; ignore what doesn’t apply.
Confidence routing
The answer says what; confidence says whether to act.
Composite scoring
Several Scores, normalised, weighted in your code.
Intent routing
Jev classifies; code, an LLM or a person handles it.
Where jev-1.13 breaks
Run them →Tricks
- 01Pair a Choice with a Noul to allow “none of these”.
- 02Combine confidences with min, not a product.
- 03Let Jev pick from candidates your code found.
- 04Cache answers; re-tune thresholds for free.
- 05Point at state fields with `ticket.messages[0].text`.
The raw requestshowhide
POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY
{
"state": "My card was charged twice.",
"model": "jev-latest",
"questions": {
"department": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": { "billing": "Charges and refunds",
"technical": "Bugs and integrations",
"other": "None of the above" }
},
"is_urgent": { "type": "noul", "instructions": "Does this convey urgency?" }
}
}