Skip to content

🐝 The First Self-Improving agents with RL / Prompting Optimization

3.4 KB
Raw Blame
from swarm.llm.format import Message
from swarm.graph import Node
from typing import List, Any, Optional
from swarm.environment.prompt.prompt_set_registry import PromptSetRegistry
from swarm.llm.format import Message
from swarm.llm import LLMRegistry


class CoTStep(Node): 
    def __init__(self, 
                 domain: str,
                 model_name: Optional[str],
                 is_last_step: bool,
                 operation_description: str = "Make one step of CoT",
                 id=None):
        super().__init__(operation_description, id, True)
        self.domain = domain
        self.model_name = model_name
        self.is_last_step = is_last_step
        self.llm = LLMRegistry.get(model_name)
        self.prompt_set = PromptSetRegistry.get(domain)
        self.role = self.prompt_set.get_role()
        self.constraint = self.prompt_set.get_constraint()

    @property
    def node_name(self):
        return self.__class__.__name__

    async def _execute(self, inputs: List[Any] = [], **kwargs):
        
        node_inputs = self.process_input(inputs)
        outputs = []
        for input_dict in node_inputs:

            role = self.prompt_set.get_role()
            constraint = self.prompt_set.get_constraint()
            if self.is_last_step:
                system_prompt = (
                    f"You are {role}. {constraint}. "
                    "Answer taking into consideration the provided sequence "
                    "of thoughts on the question at hand.")
            else:
                system_prompt = (
                    f"You are {role}. "
                    "Given the question, solve it step by step. "
                    "Answer your thoughts about the next step of the solution given "
                    "everything that has been provided to you so far. "
                    "Expand on the next step. "
                    "Do not try to provide the answer straight away, instead expand "
                    "on your thoughts about the next step of the solution."
                    "Aswer in maximum 30 words. "
                    "Do not expect additional input. Make best use of whatever "
                    "knowledge you have been already provided.")
            if 'output' in input_dict:
                task = input_dict['output']
            else:
                task = input_dict["task"]
            user_prompt = self.prompt_set.get_answer_prompt(question=task)
            message = [
                Message(role="system", content=system_prompt),
                Message(role="user", content=user_prompt)]
            response = await self.llm.agen(message, max_tokens=50)
            if self.is_last_step:
                concatenated_response = response
            else:
                concatenated_response = f"{task}. Here is the next thought. {response}. "

            execution = {
                "operation": self.node_name,
                "task": task,
                "files": input_dict.get("files", []),
                "input": task,
                "role": role,
                "constraint": constraint,
                "prompt": user_prompt,
                "output": concatenated_response,
                "ground_truth": input_dict.get("GT", []),
                "format": "natural language"
            }
            outputs.append(execution)
            self.memory.add(self.id, execution)
            # self.log()

        return outputs