strands.decisions.llm_decision_model

Decision model backed by a vended Model provider.

Translates an ask call into a single forced structured-output model call, parses the model’s answers, and lifts them back into the SDK’s answer shape.

LLMDecisionModel

class LLMDecisionModel(DecisionModel)

Defined in: src/strands/decisions/llm_decision_model.py:31

Adapter that turns any Model into a DecisionModel.

ask is implemented as one direct model call with a dynamically-built Pydantic answer model and forced structured output. Uncertainty is modeled by appending an 'uncertain' option to each permitting question’s enum.

Example:

decision = LLMDecisionModel(model=BedrockModel())
result = await decision.ask(
    "The fruit is an apple",
    \{
        "color": \{"question": "What color is the fruit?", "choices": ["red", "green", "blue"]},
        "is_fruit": \{
            "question": "Is the thing talked about a fruit?",
            "choices": "boolean",
            "uncertainty": False,
        },
    },
)
# result.answers["color"]    is "red" | "green" | "blue" | Uncertain
# result.answers["is_fruit"] is True | False | Uncertain
#   (opting out via `uncertainty=False` only removes 'uncertain' from the model's enum)

__init__

def __init__(model: Model | None = None,
             *,
             system_prompt: str = _DEFAULT_SYSTEM_PROMPT) -> None

Defined in: src/strands/decisions/llm_decision_model.py:58

Create an LLM-backed decision model.

Arguments:

• model - Underlying vended model. Defaults to a zero-config BedrockModel.
• system_prompt - System prompt used to frame the decision for the model. The SDK appends mandatory output and prompt-injection rules that the prompt cannot override. Defaults to a generic decision framing.

Raises:

• TypeError - If model is not a Model.