import evoagentx.workflow.operators as operator
import examples.aflow.pertqa.optimized_adamson_update.round_18.prompt as prompt_custom
from evoagentx.models.model_configs import LLMConfig
from evoagentx.benchmark.benchmark import Benchmark
from evoagentx.models.model_utils import create_llm_instance

class Workflow:
    
    def __init__(
        self,
        name: str,
        llm_config: LLMConfig,
        benchmark: Benchmark
    ):
        self.name = name
        self.llm = create_llm_instance(llm_config)
        self.benchmark = benchmark
        self.custom = operator.Custom(self.llm)
        self.answer_generate = operator.AnswerGenerate(self.llm)
        self.qas_ensemble = operator.QAScEnsemble(self.llm)
    
    async def __call__(self, problem: str):
        """
        Implementation of the workflow
        """
        solutions_list = []
        for _ in range(3):  # Collecting multiple answers
            solution = await self.answer_generate(input=problem)
            solutions_list.append(solution['answer'])
        
        review_responses = []
        for answer in solutions_list:  # Review each generated answer
            review_response = await self.custom(input=problem + " Review the answer: " + answer, instruction=prompt_custom.Review_PROMPT)
            review_responses.append(review_response['response'])
        
        ensemble_response = await self.qas_ensemble(solutions=review_responses)
        
        # Added self-consistency check before final review
        if ensemble_response['response'] not in solutions_list:
            consistency_check = await self.custom(input=problem + " Check consistency of the ensemble answer: " + ensemble_response['response'], instruction=prompt_custom.Review_PROMPT)
            solutions_list.append(consistency_check['response'])
        
        final_review_response = await self.custom(input=problem + " Final review of the ensemble answer: " + ensemble_response['response'], instruction=prompt_custom.Review_PROMPT)
        return final_review_response['response']  # Return the final reviewed response