Behaviour for stream chunk normalization.

Implementations transform provider-specific stream chunks into normalized Nous stream events:

• {:text_delta, text} - Incremental text content
• {:thinking_delta, text} - Incremental reasoning/thinking content
• {:tool_call_delta, tool_calls} - Tool call information
• {:finish, reason} - Stream completion signal
• {:unknown, chunk} - Unrecognized chunk (filtered by default)

Default Implementation

Most OpenAI-compatible providers work with the default implementation:

Nous.StreamNormalizer.OpenAI

Custom Implementation

For providers with unique formats, implement this behaviour:

defmodule MyApp.CustomNormalizer do
  @behaviour Nous.StreamNormalizer

  @impl true
  def normalize_chunk(chunk) do
    # Your custom logic
    [{:text_delta, chunk["custom_field"]}]
  end

  @impl true
  def complete_response?(chunk), do: false

  @impl true
  def convert_complete_response(_chunk), do: []
end

Then configure it on the model:

Nous.new("openai_compatible:custom-model",
  base_url: "http://custom-server/v1",
  stream_normalizer: MyApp.CustomNormalizer
)