Nullability not respected in JSON Schema generation

[ponychicken]

Right now in the generated json schema all fields are marked as required (i think its a good trick to force the model to populate everything) but the types of fields are never allowed to be null. Even though the Elixir schema would accept a null value fine.
For complex documents there might be fields which are not always present in the source text, and forcing the model to put something which is not present just results in hallucinations.

[daveb1976]

I'd second this. In the mean time I managed to work around the issue by explicitly instructing the llm to provide nil values for the fields.

defmodule UserProfile do
  use Ecto.Schema
  use Instructor

  @llm_doc """
  ## Field Descriptions:
  - bio: A brief biography of the user.
  - interests: A list of user's interests.

  Note: The 'name' and 'age' fields will be provided separately, so please produce nil values for these.
  """
  @primary_key false
  embedded_schema do
    field(:name, :string, default: nil)
    field(:age, :integer, default: nil)
    field(:bio, :string)
    field(:interests, {:array, :string})
  end

  def validate_changeset(changeset) do
    changeset
    |> Ecto.Changeset.validate_required([:bio, :interests])
    |> Ecto.Changeset.validate_length(:bio, max: 500)
    |> Ecto.Changeset.validate_length(:interests, min: 1)
  end
end

then

# Generate LLM data
{:ok, llm_data} =
  Instructor.chat_completion(
    model: "gpt-4o",
    response_model: UserProfile,
    max_retries: 3,
    messages: [
      %{
        role: "user",
        content: """
        Generate a brief bio and list of interests for a user.
        """
      }
    ]
  )

result in

{:ok,
 %UserProfile{
   name: nil,
   age: nil,
   bio: "An avid traveler and food enthusiast, always seeking new adventures and culinary experiences around the world.",
   interests: ["traveling", "cooking", "photography", "reading", "hiking"]
 }}