Tool Selector

Select relevant regular tools before language-model invocation.

DMToolSelector(dm_invoker, threshold=_DEFAULT_THRESHOLD, relevance_question=None)

Bases: BaseToolSelector

[BETA] Select regular tools using one binary relevance question per candidate.

The complete candidate catalog is sent in one decisions-model request. Selection is a filtering optimization only; callers must authorize candidate tools before invoking this selector.

Examples:

dm_invoker = build_dm_invoker(...)
tool_selector = DMToolSelector(dm_invoker)
selected_tools = await tool_selector.select(messages=messages, tools=tools)

Attributes:

Name	Type	Description
dm_invoker	BaseDMInvoker	The injected decisions-model invoker.
threshold	float	Inclusive probability threshold for retaining a tool.
relevance_question	DMQuestion	Custom noul question definition used for every candidate tool.

Initialize a decisions-model tool selector.

Parameters:

Name	Type	Description	Default
dm_invoker	BaseDMInvoker	The injected decisions-model invoker.	required
threshold	float	Inclusive probability threshold for retaining a tool. Defaults to _DEFAULT_THRESHOLD.	_DEFAULT_THRESHOLD
relevance_question	DMQuestion | None	Custom noul question definition used for every candidate tool. Defaults to None, in which case the built-in question is used.	None

Raises:

Type	Description
InferenceValueError	If threshold is not a finite probability in [0, 1], or relevance_question is not a noul question.

select(messages, tools, *, max_tools=None) async

Return candidate tools that the decisions model considers relevant.

Parameters:

Name	Type	Description	Default
messages	list[Message]	Complete canonical messages for the LM request.	required
tools	list[Tool]	Already-authorized candidate tools with unique names.	required
max_tools	int | None	Positive cap on selected tools. Highest-probability tools are retained, then returned in original input order. Defaults to None.	None

Returns:

Type	Description
list[Tool]	list[Tool]: Original selected tools in input order.