Skip to content

datarobot_genai.langgraph.reasoning

reasoning

LangChain-message reasoning normalization for the LangGraph adapter.

Wraps the framework-agnostic iter_content_blocks (in core) with the LangChain BaseMessage surface so it can type the message directly instead of duck-typing. Lives in the langgraph package because it depends on LangChain; core stays framework-free.

iter_message_blocks 

iter_message_blocks(message: BaseMessage) -> Iterator[tuple[Literal['text', 'thinking'], str]]

Yield typed (kind, delta) pairs for any AIMessage/AIMessageChunk shape.

Combines two surfaces LangChain/LiteLLM produce for reasoning models:

  • Native list-form content: e.g. Anthropic/Bedrock blocks [{"type": "thinking", ...}, {"type": "text", ...}] — delegated to iter_content_blocks(message.content).
  • OpenAI-compatible flat shape: text in message.content (string) and reasoning hoisted to message.additional_kwargs["reasoning_content"] (string). This is what the DataRobot LLM gateway returns over its OpenAI-style HTTP API, even when the underlying model is Anthropic.

These shapes are not mutually exclusive: with extended thinking enabled the DataRobot gateway emits the same reasoning delta in BOTH additional_kwargs["reasoning_content"] and as a native content thinking block. To avoid double-emitting, the flat reasoning_content is only used as a fallback when content does not already carry thinking.

Reasoning is yielded before content so consumers can route REASONING_* events ahead of the matching text in the AG-UI stream.

Source code in datarobot_genai/langgraph/reasoning.py 
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
def iter_message_blocks(
    message: BaseMessage,
) -> Iterator[tuple[Literal["text", "thinking"], str]]:
    """Yield typed (kind, delta) pairs for any AIMessage/AIMessageChunk shape.

    Combines two surfaces LangChain/LiteLLM produce for reasoning models:

    - **Native list-form content**: e.g. Anthropic/Bedrock blocks
      ``[{"type": "thinking", ...}, {"type": "text", ...}]`` — delegated to
      ``iter_content_blocks(message.content)``.
    - **OpenAI-compatible flat shape**: text in ``message.content`` (string) and
      reasoning hoisted to ``message.additional_kwargs["reasoning_content"]``
      (string). This is what the DataRobot LLM gateway returns over its
      OpenAI-style HTTP API, even when the underlying model is Anthropic.

    These shapes are not mutually exclusive: with extended thinking enabled the
    DataRobot gateway emits the *same* reasoning delta in BOTH
    ``additional_kwargs["reasoning_content"]`` and as a native ``content``
    thinking block. To avoid double-emitting, the flat ``reasoning_content`` is
    only used as a fallback when ``content`` does not already carry thinking.

    Reasoning is yielded before content so consumers can route REASONING_*
    events ahead of the matching text in the AG-UI stream.
    """
    content_pairs = list(iter_content_blocks(message.content))
    reasoning_text = message.additional_kwargs.get("reasoning_content")
    if reasoning_text and not any(kind == "thinking" for kind, _ in content_pairs):
        yield "thinking", reasoning_text
    yield from content_pairs