Skip to main content
The Torchcast Decision API uses the standard decisions API request, response and error shapes for POST /v1/decisions. Most existing integrations move with three changes.

​
What to change

  1. Base URL: send requests to https://api.torchcast.ai/v1/decisions.
  2. API key: create a key in the console and send it as Authorization: Bearer tcd_....
  3. Model id: use torchcast-decider-latest, or a pinned id from GET /v1/models.
cURL
Your existing parsing of answers, usage and error keeps working.

​
Rules to check

​
The response model is the pinned id

The model field of a successful response is always the pinned id that answered, never the alias you sent. If you call torchcast-decider-latest, expect model to be a pinned id such as torchcast-decider-4b-v1-1-api-1. If your code compares the response model with the requested one, compare against the alias’s target_model from GET /v1/models instead, or drop the check. See Models.

​
Question names

Question names must be 1 to 64 characters of letters, digits, _ and -. Names with spaces, dots or other characters are rejected with 400 invalid_request. Rename such questions, for example "is defect?" to is_defect.

​
Option names

choice option names must be 1 to 128 characters, and each must be unique within its question. A JSON body that repeats a key anywhere (in questions, in criteria, or inside the state) is rejected with 400, rather than silently keeping the last value.

​
Strict request fields

Fields other than model, state and questions at the top level, and other than type, instructions and criteria in a question, are rejected with 400 invalid_request, with param naming the field.

​
Body size

Request bodies can be up to 28 MiB. See Limits.

​
Other behavior

  • Validation errors are always 400, never 422.
  • x-request-id is generated by Torchcast on every response; a value you send is ignored.
  • There is no Idempotency-Key support: a repeated successful request is charged again. See Limits.
  • Prices and per-model limits come from GET /v1/models. See Billing.