nat.experimental.test_time_compute.search.multi_query_retrieval_search

Attributes

logger

Classes

MultiQueryRetrievalSearch

A strategy that, for each incoming TTCItem, generates multiple new items by

Functions

register_multi_query_retrieval_search(config, builder)

Module Contents

logger

class MultiQueryRetrievalSearch(

config: nat.experimental.test_time_compute.models.search_config.MultiQueryRetrievalSearchConfig,

)

Bases: nat.experimental.test_time_compute.models.strategy_base.StrategyBase

A strategy that, for each incoming TTCItem, generates multiple new items by re-writing the input ‘task_description’ from different perspectives. Uses multiple LLMs to encourage diversity.

config

llms_bound = []

async build_components(builder: nat.builder.builder.Builder) → None

Binds each LLMRef in self.config.llms to an actual LLM client.

supported_pipeline_types() → list[nat.experimental.test_time_compute.models.stage_enums.PipelineTypeEnum]

Return the stage types supported by this selector.

stage_type() → nat.experimental.test_time_compute.models.stage_enums.StageTypeEnum

Return the stage type of this strategy.

async ainvoke(items: list[nat.experimental.test_time_compute.models.ttc_item.TTCItem], original_prompt: str | None = None, agent_context: str | None = None, \*\*kwargs) → list[nat.experimental.test_time_compute.models.ttc_item.TTCItem]

For each TTCItem, rewrite the ‘input’ using each LLM to create a new perspective. The new TTCItems’ ‘output’ field will store the newly generated query.

async register_multi_query_retrieval_search(

config: nat.experimental.test_time_compute.models.search_config.MultiQueryRetrievalSearchConfig,

builder: nat.builder.builder.Builder,

)