What answers the questions: a behaviour with one callback.

Jev.HTTP is the backend that speaks to TypeSafe and compatible servers. An in-process model, a cache, or a fake for tests is another module with the same shape. jev_nx is one: it runs open decision models such as Laya on Nx, in your own application.

defmodule Canned do
  @behaviour Jev.Backend

  @impl true
  def post(state, questions, opts) do
    Jev.Telemetry.span(state, questions, %{backend: __MODULE__, tag: opts[:tag]}, fn ->
      {:ok, Jev.reply(Jev.Test.body([kind: :bug], questions), questions)}
    end)
  end
end

A Jev.Server picks the backend per request or from configuration:

{:reply, {tag, state, [kind: @kinds], [backend: Canned]}, s}

config :jev, backend: Canned

The callback receives the state, the questions as Jev.questions/1 returns them, and the request options with tag among them. It returns {:ok, reply} with the map Jev.reply/3 builds, or {:error, reason}. Wrapping the work in Jev.Telemetry.span/4 is what makes the cost and calibration telemetry, and the dashboards on it, work for every backend.