Decisions

POST /v1/decisions asks a model several typed questions about one piece of state in a single call. Each named question gets its own answer: a probability, a chosen option, or a score. Use it for the narrow judgments described in Decision models, such as routing, moderation, and guards, through the same endpoint and virtual keys as the rest of your model traffic.

Call the endpoint

        
    
curl https://<your-gateway-endpoint>/v1/decisions \
  -H "Authorization: Bearer sk-un-..." \
  -H "Content-Type: application/json" \
  -d '{
  "model": "typesafe/<model>",
  "state": "Customer: I was charged twice for one order.",
  "questions": {
    "is_billing": {
      "kind": "noul",
      "instructions": "Is this a billing issue?"
    },
    "urgency": {
      "kind": "choice",
      "instructions": "How urgent is it?",
      "criteria": {
        "low": "can wait",
        "high": "reply today"
      }
    }
  }
}'

If the gateway is behind Union sign-in, run flyte proxy app first and call http://127.0.0.1:8600/v1/decisions instead. See Use from your laptop.

Replace <your-gateway-endpoint> with your gateway’s endpoint from Settings → Inference Gateway, sk-un-... with a virtual key, and <model> with a model your typesafe provider allows.

When the gateway has a deployed typesafe provider, or a vLLM or SGLang provider with Answers decisions on, the gateway’s overview in Settings → Inference Gateway shows this command with your endpoint and a model filled in.

The request

Field	Meaning
`model`	The `provider/model` that answers, for example `typesafe/<model>` or `qwen/qwen3-8b`. The provider must answer decisions; see Which providers answer decisions.
`state`	The text the questions are about.
`questions`	The questions, keyed by name.

Each question has a kind and instructions, the question itself:

`kind`	Answer	`criteria`
`noul`	A probability.	Not needed.
`choice`	One of the options.	A mapping of each option to a description of when it applies.

Which providers answer decisions

Provider	Answers decisions
`typesafe`	Yes. TypeSafe’s decision models answer decisions only, not chat or completions.
Hosted LLM providers	Chat models that support tool calling. The gateway asks the model through one forced tool call.
vLLM and SGLang apps	Only when the provider was added with Answers decisions on. The gateway asks the model through one forced tool call on chat completions, so the model must support tool calling. With the setting off, the gateway refuses decision requests for that provider. See Add a self-hosted model.

A gateway on an older image may not serve /v1/decisions. If the gateway’s overview shows an update available, redeploy the gateway.