📘 Overview

The AtomTaskGenerator is an operator that generates an atomic task based on a given input scenario. It utilizes a Large Language Model (LLM) service to understand the scenario and formulate a corresponding atomic task.

__init__ function

def __init__(self, llm_serving: LLMServingABC)

Parameter	Type	Default	Description
llm_serving	LLMServingABC	Required	LLM serving object that implements the LLMServingABC interface.

Prompt Template Descriptions

Prompt Template Name	Primary Use	Applicable Scenarios	Feature Description

run function

def run(self, storage: DataFlowStorage, input_scenario_key: str, output_key: str = "atom_task")

Parameter	Type	Default	Description
storage	DataFlowStorage	Required	DataFlow storage instance, responsible for reading and writing data.
input_scenario_key	str	Required	Input column name, corresponding to the scenario field.
output_key	str	"atom_task"	Output column name, corresponding to the generated atomic task field.

🧠 Example Usage

from dataflow.operators.conversations import AtomTaskGenerator
from dataflow.utils.storage import FileStorage
from dataflow.serving import APILLMServing_request
from dataflow.core import LLMServingABC

class AtomTaskGeneratorExample:
    def __init__(self, llm_serving: LLMServingABC = None):
        self.storage = FileStorage(
            first_entry_file_name="input.jsonl",
            cache_path="./cache_local",
            file_name_prefix="dataflow_cache_step",
            cache_type="jsonl",
        )

        self.llm_serving = APILLMServing_request(
            api_url="",
            model_name="gpt-4o",
            max_workers=30
        )

        self.generator = AtomTaskGenerator(
            llm_serving=self.llm_serving
        )

    def forward(self):
        self.generator.run(
            storage=self.storage.step(),
            input_scenario_key="scenario",
            output_key="atom_task"
        )

if __name__ == "__main__":
    pl = AtomTaskGeneratorExample()
    pl.forward()

🧾 Default Output Format

Field	Type	Description
input_scenario_key	str	The input scenario text. The actual field name is determined by the input_scenario_key parameter.
atom_task	str	The generated atomic task. The actual field name is determined by the output_key parameter.

Example Input: (Assuming input_scenario_key is "scenario")

{
  "scenario": "Attempting to book a flight."
}

Example Output:

{
  "scenario": "Attempting to book a flight.",
  "atom_task": "Search for a round-trip flight from New York to London, departing on November 15th and returning on November 22nd, with a budget of $500 or less."
}