Decisions

POST /v1/decidePOST /v1/decide/batchPOST /v1/systemonebilled per decision

Give the model a piece of state (a message, ticket, call transcript or form record) and typed questions about it. It returns a calibrated probability for every option, and works in English and eight African languages. Powered by PurpleMIST-Flash-1.0.

result = call("/v1/decide", {
    "state": {"channel": "whatsapp",
              "message": "Ẹ jọ̀ọ́, wọ́n ti yọ owó lẹ́ẹ̀mejì lórí káàdì mi. Mo fẹ́ kí ẹ dá owó mi padà."},
    "questions": {
        "team":    {"type": "choice", "instructions": "Which team should handle this?",
                    "criteria": {"billing": "payments and refunds", "delivery": "orders in transit"}},
        "refund":  {"type": "noul", "instructions": "The customer is asking for money back."},
        "urgency": {"type": "score", "instructions": "How urgent is this?",
                    "criteria": ["can wait", "normal queue", "today", "immediately"]},
    },
})

answers = result["answers"]
if answers["refund"]["noul"] > 0.9:
    route_to(answers["team"]["choice"])

curl -X POST "$OLAVERSE_URL/v1/decide" \
  -H "Authorization: Bearer $OLAVERSE_KEY" -H "Content-Type: application/json" \
  -d '{"state": "I was charged twice, please refund me",
       "questions": {"refund": {"type": "noul", "instructions": "The customer asks for money back."}}}'

Request

Field	Type	Notes
`model`	string	Optional. `purplemist-flash-1.0` (default).
`state`	string, object or array	The thing being decided about. It must not be empty.
`questions`	object	1 to 64 questions, keyed by your own IDs.

Question types

`type`	`criteria`	You get back
`choice`	2 to 255 option keys, each mapped to a description (or `null`)	`choice` (best key), `probabilities`, `confidence`
`noul`	Optional; only `true` / `false` keys	`noul`: the probability the statement is true
`score`	A list of 2 to 26 level descriptions, lowest first	`score` (expected level), `probabilities`, `legend`, `confidence`

Response

Values are illustrative.

{
  "model": "purplemist-flash-1.0",
  "answers": {
    "team":    {"type": "choice", "choice": "billing",
                "probabilities": {"billing": 0.98, "delivery": 0.02}, "confidence": 0.86},
    "refund":  {"type": "noul", "noul": 0.97},
    "urgency": {"type": "score", "score": 1.9,
                "probabilities": {"0": 0.05, "1": 0.25, "2": 0.45, "3": 0.25},
                "legend": {"0": "can wait", "1": "normal queue", "2": "today", "3": "immediately"},
                "confidence": 0.12}
  },
  "usage": {"decisions": 3}
}

Reading the answers

Probabilities are calibrated, so thresholds mean what they say. For example, act automatically above 0.9 and send the rest to a person.
score is the expected level. 1.9 means “close to today”. For the single most likely level, take the highest entry in probabilities.
confidence measures how peaked the distribution is (0 = no idea, 1 = certain). It is not a second probability.
Ask several questions at once. They share one pass over the state, so they cost no extra time.
Long inputs are never truncated. A state of up to about 32,000 tokens is read whole. Anything longer returns 413 with a “maximum context length” message, so split it into separate calls.

Batch

POST /v1/decide/batch takes up to 32 cases per call. Give each case an id to match the results back:

result = call("/v1/decide/batch", {
    "cases": [
        {"id": "t-1001", "state": "Refund me now!", "questions": {"refund": {"type": "noul", "instructions": "Asks for a refund."}}},
        {"id": "t-1002", "state": "Where is my parcel?", "questions": {"refund": {"type": "noul", "instructions": "Asks for a refund."}}},
    ],
})
for r in result["results"]:
    print(r["id"], r.get("answers") or r["error"])

Each case succeeds or fails on its own, so one bad case doesn’t sink the batch. Failed cases come back as {"id", "error", "status"} (400 invalid, 413 too long), and only answered questions are billed.