If you know the list of call groups to route to at compile time you could use Instructor and do something like this:

defmodule CallRouter do
  use Ecto.Schema
  
  @doc """
    ## Field Descriptions:
    - call_group: Call group to route to
  """
  @primary_key false
  embedded_schema do
    field(:call_group, Ecto.Enum, values: [:group_1, :group_2, ..., :group_n])
  end

  def route_call(data) do
    Instructor.chat_completion(
    model: "gpt-3.5-turbo",
    response_model: __MODULE__,
    max_retries: 3,
    messages: [
        %{
          role: "system",
          content: """
          Your job is to route calls based on the given data ....
          """
      },
    %{
        role: "user",
        content: """
        Here is the data to consider: #{inspect(data)}
        """
      }
    ]
  )
  end
end

Then to use it you would do something like:

case CallRouter.route_call(data) do
  {:ok, %CallRouter{call_group: group} = result} ->
    # do something with group
  {:error, %Ecto.Changeset = changeset} ->
    # handle errors
end

Right now the adapters supported by Instructor are OpenAI, llamacpp, and (I think) ollama, and bumblebee is on their roadmap, but I dont think it’s supported quite yet.