Sikaru decisions (System One)
System One answers named questions about one state, which is a text or a JSON value. There are three question types:
| Type | Use for | Shape |
|---|---|---|
noul |
Yes or no | an instruction |
choice |
Pick one label | criteria maps each label to null (or a description) |
score |
An ordered scale | criteria lists levels from low to high |
A noul answer is the probability of yes, from 0 to 1. A choice or score answer
comes with per-option probabilities and a confidence. Use these numbers: route on a
threshold, or send uncertain cases to a human. Don't treat every answer as certain.
If the model couldn't answer a question, that question is missing from answers, so
check for it. Each question can have at most 26 options.
From the CLI
# Yes-or-no questions inline (repeatable).
sikaru decision infer "I was charged twice. Please fix this ASAP." \
--noul urgent="Is this urgent?" --noul refund="Does the customer want a refund?"
# Choice and score questions from a file, as JSON.
cat > questions.json <<'JSON'
{
"category": {"type": "choice", "instructions": "What is this ticket about?",
"criteria": {"billing": null, "technical": null, "other": null}},
"severity": {"type": "score", "instructions": "How severe is it?",
"criteria": ["low", "medium", "high"]}
}
JSON
sikaru decision infer "I was charged twice." --questions questions.json -o json
# A JSON state from stdin.
cat ticket.json | sikaru decision infer --json --questions questions.json -o json
Signed-in calls bill the saved project.
From code
import os
from sikaru_sdk import SikaruApi
client = SikaruApi(api_key=os.environ["SIKARU_API_KEY"]) # defaults to https://api.sikaru.ai
decision = client.decisions.create(
state="I was charged twice.",
questions={"refund": {"type": "noul", "instructions": "Does the customer want a refund?"}},
)
TypeScript has the same call, client.decisions.create(...), on SikaruApi from
@aadi-labs/sikaru-sdk. Check exact fields with sikaru decision infer --schema or the OpenAPI
schema (POST /v1/systemone).
Writing good questions
- Make each question about one thing, and name it for the field your code reads.
- Make choice labels mutually exclusive. Add an
otherlabel when the set isn't closed. - Order score levels from low to high, and keep them few.
- Put only the facts needed for the decision in the state. It can contain customer data, so don't log it carelessly.