Decisions API

POST /v1/decisions sends the decision model a piece of context (state) and one or more typed questions, and returns the model's answers. It is built for classification and routing: triage a support message, pick a category, score a probability — decisions rather than prose.

This is a separate surface from chat: the decision model is not a chat model, it is deliberately not listed in GET /v1/models, and /v1/decisions is the only way to reach it.

#Quickstart

bash

curl https://api.radienthq.com/v1/decisions \
  -H "Authorization: Bearer $RADIENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Customer writes: I was charged twice for my invoice this month and want a refund.",
    "questions": {
      "topic": {
        "type": "choice",
        "instructions": "Which department should handle this message?",
        "criteria": {
          "billing": "Payments, invoices, refunds",
          "technical": "Bugs and outages",
          "other": "Anything else"
        }
      }
    }
  }'

json

{
  "answers": {
    "topic": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 1, "technical": 0, "other": 0 },
      "confidence": 1
    }
  },
  "id": "gen-dec-…",
  "model": "…",
  "provider": "…",
  "usage": { "input_tokens": 345, "output_tokens": 38, "cost": 0.00001449 }
}

The model and provider values above are trimmed: they identify the model and the serving endpoint. The numbers are as returned.

#Request

Field	Type	Required	Notes
`state`	string	yes	The context the questions are asked about. Put the full material here; there is no separate prompt field.
`questions`	object	yes	One entry per question, keyed by the id you want back in `answers`.
`model`	string	no	Defaults to `jev-1.13`. Other configured spellings are mapped to the model the deployment serves.

Each question object:

Field	Type	Required	Notes
`type`	string	yes	`choice` for a selection from a fixed set; `noul` for a probability between 0 and 1.
`instructions`	string	yes	What the model should decide.
`criteria`	object	for `choice`	`{label: description}` — the allowed answers. Answers come back as one of the labels (a string).

The request body is capped at 1 MiB; a larger body is refused with 413 {"error":"The decision request body is too large"}.

#Response

answers — keyed by your question ids. A choice answer carries choice (one of your labels), probabilities (one per label) and confidence; a noul answer carries noul, a probability between 0 and 1:

json

"is_urgent": { "type": "noul", "noul": 0.99 }

id, model, provider — the request id, the model that answered, and the serving endpoint.
usage — input_tokens, output_tokens, and cost in USD.

The answer body comes from the decision model, so additional fields may appear as it evolves. Treat unknown fields as informational.

#Errors

Status	Body	Cause
`400`	`{"error":"Could not read the request body"}`	The request body could not be read (from the controller; not reproducible with a stock client)
`401`	auth bodies	No or invalid credential (Authentication)
`402`	`{"error":"insufficient credits"}`	The credit gate refused the call
`405`	`{"error":"Method Not Allowed"}`	`GET /v1/decisions` — the route is POST-only
`413`	`{"error":"The decision request body is too large"}`	Body over 1 MiB
`422`	the model's own body, naming the offending field	The model itself rejected the request; its body is passed through as returned (from the controller's passthrough set; not reproducible with a stock model)
`429`	the model's own body	Rate limited by the model; `Retry-After` is passed through when present
`502`	`{"error":"The decision model upstream failed with status N"}`	Any upstream status other than 2xx, `422`, `429` or `529`, with that status in the message
`502`	`{"error":"The decision model upstream could not be reached"}`	The model service could not be reached at all
`503`	a `Decisions are unavailable…` body	This deployment has no decision model configured. Nothing was metered or billed.

Only the model's own 2xx, 422, 429 and its overloaded 529 are passed through with the model's body. Every other status — including the model's own 400-class rejections — becomes 502 {"error":"The decision model upstream failed with status N"}; a non-JSON body, an empty body and an unusable question shape were each executed and measured this way.

#Cost and limits

Priced per input token: $0.042 per million tokens ($0.000000042 per token). Output tokens appear in usage but are not charged. Example: a 345-input-token call costs $0.00001449.
The credit gate reserves up to $0.30 per call before admitting it (see the credit gate).
An upstream "overloaded" status (529) is passed through with the model's body.
The model call has a 30-second server-side timeout; there are no server-side retries or failover.
There is no request-rate limit on this route beyond the credit gate and the model's own 429s.

See also: Usage and credits API · API Reference.