Elixir client for the TypeSafe AI API.

TypeSafe evaluates typed questions against a state and returns structured answers your code can branch on directly — no text parsing required.

Question types

Quickstart

# Configure once (reads TYPESAFE_API_KEY from env by default)
client = ExTypesafe.Client.new()

# Build typed questions. Atom keys are retained in the response.
questions = %{
  is_urgent: ExTypesafe.Question.noul("Does this convey urgency?"),
  department: ExTypesafe.Question.choice(
    "Which team should handle this?",
    %{billing: "Payments, invoicing, refunds", technical: "Bugs, outages, integrations"}
  ),
  frustration: ExTypesafe.Question.score(
    "How frustrated is the customer?",
    ["Calm", "Frustrated", "Very angry"]
  )
}

state = "Hi, my payouts have been failing for 3 days. I'm losing sales. Please help!"

case ExTypesafe.system_one(client, state, questions) do
  {:ok, response} ->
    response.answers.is_urgent.noul       #=> 0.97
    response.answers.department.choice    #=> "billing"
    response.answers.frustration.score    #=> 1.83
    response.scores.frustration.legend    #=> %{"0" => "Calm", ...}

  {:error, error} ->
    IO.inspect(error)
end

Configuration

See ExTypesafe.Client.new/1 and ExTypesafe.Config for all options.

Environment variables are supported:

Retries

429 Too Many Requests, 529 Overloaded, and transport-level connection failures are retried automatically with capped exponential backoff. Numeric Retry-After (delta seconds) and retry-after-ms headers are honored when within the configured cap. Configure :max_retries, :retry_delay_ms, and :max_retry_delay_ms on the client or per call. A transport retry can repeat a POST if the connection fails after the server receives it.