Kedi Laya

Local Laya decision models without a TypeSafe dependency. Pydantic AI and LangChain are optional integrations, not required to import the client.

Use kedi-decisions[laya-mlx] on Apple Silicon or kedi-decisions[laya] for upstream PyTorch. Add langchain for that framework. The current packages are development versions in this monorepo and have not been published yet.

from kedi_laya import LayaClient

with LayaClient("path/to/checkpoint", backend="mlx") as client:
    evaluator = client.as_evaluator(threshold=0.85)
    result = evaluator.evaluate_sync(
        state="The duplicate charge has been refunded.",
        schema={
            "type": "object",
            "properties": {"refunded": {"type": "boolean", "description": "Has the refund been completed?"}},
            "required": ["refunded"],
        },
    )
    print(result.values)

LayaModel and LayaChatModel are available from kedi_laya or their respective integrations.pydantic and integrations.langchain modules. They accept a checkpoint, backend, threshold, and load_options. A borrowed client can be shared between models; closing a wrapper never closes a borrowed client.

backend="auto" prefers an installed MLX runtime on Apple Silicon, otherwise selects Torch. Explicit mlx or torch makes deployment deterministic. Missing or failing selected runtimes raise an error; there is no inference-error fallback to another backend. Weights load lazily once per client. The provider identifier is always laya, and model.backend reports the selected runtime.

MLX-owned models are checked against the actual tokenizer and context budget before inference. Injected predictors and Torch retain their own context handling. Real inference has been verified on the multilingual MLX checkpoint; Torch must not be considered verified by the mocked loader tests.

The optional tool selector only executes parameterless calls through the host framework. Parameterized tool choices raise ToolCallProposed; the model does not invent arguments. Raw generated text and unsupported schemas are rejected.