NexQuant/rdagent/scenarios/qlib/developer/feedback.py

import json
from pathlib import Path
from typing import Dict

import pandas as pd
from jinja2 import Environment, StrictUndefined

from rdagent.core.experiment import Experiment
from rdagent.core.prompts import Prompts
from rdagent.core.proposal import (
    Experiment2Feedback,
    Hypothesis,
    HypothesisFeedback,
    Trace,
)
from rdagent.log import rdagent_logger as logger
from rdagent.oai.llm_utils import APIBackend
from rdagent.utils import convert2bool

feedback_prompts = Prompts(file_path=Path(__file__).parent.parent / "prompts.yaml")
DIRNAME = Path(__file__).absolute().resolve().parent


def process_results(current_result, sota_result):
    # Convert the results to dataframes
    current_df = pd.DataFrame(current_result)
    sota_df = pd.DataFrame(sota_result)

    # Set the metric as the index
    current_df.index.name = "metric"
    sota_df.index.name = "metric"

    # Rename the value column to reflect the result type
    current_df.rename(columns={"0": "Current Result"}, inplace=True)
    sota_df.rename(columns={"0": "SOTA Result"}, inplace=True)

    # Combine the dataframes on the Metric index
    combined_df = pd.concat([current_df, sota_df], axis=1)

    # Select important metrics for comparison
    important_metrics = [
        "1day.excess_return_without_cost.max_drawdown",
        "1day.excess_return_without_cost.information_ratio",
        "1day.excess_return_without_cost.annualized_return",
        "IC",
    ]

    # Filter the combined DataFrame to retain only the important metrics
    filtered_combined_df = combined_df.loc[important_metrics]

    filtered_combined_df[
        "Bigger columns name (Didn't consider the direction of the metric, you should judge it by yourself that bigger is better or smaller is better)"
    ] = filtered_combined_df.apply(
        lambda row: "Current Result" if row["Current Result"] > row["SOTA Result"] else "SOTA Result", axis=1
    )

    return filtered_combined_df.to_string()


class QlibFactorExperiment2Feedback(Experiment2Feedback):
    def generate_feedback(self, exp: Experiment, trace: Trace) -> HypothesisFeedback:
        """
        Generate feedback for the given experiment and hypothesis.

        Args:
            exp (QlibFactorExperiment): The experiment to generate feedback for.
            hypothesis (QlibFactorHypothesis): The hypothesis to generate feedback for.
            trace (Trace): The trace of the experiment.

        Returns:
            Any: The feedback generated for the given experiment and hypothesis.
        """
        hypothesis = exp.hypothesis
        logger.info("Generating feedback...")
        hypothesis_text = hypothesis.hypothesis
        current_result = exp.result
        tasks_factors = [task.get_task_information_and_implementation_result() for task in exp.sub_tasks]
        sota_result = exp.based_experiments[-1].result

        # Process the results to filter important metrics
        combined_result = process_results(current_result, sota_result)

        # Generate the system prompt
        sys_prompt = (
            Environment(undefined=StrictUndefined)
            .from_string(feedback_prompts["factor_feedback_generation"]["system"])
            .render(scenario=self.scen.get_scenario_all_desc())
        )

        # Generate the user prompt
        usr_prompt = (
            Environment(undefined=StrictUndefined)
            .from_string(feedback_prompts["factor_feedback_generation"]["user"])
            .render(
                hypothesis_text=hypothesis_text,
                task_details=tasks_factors,
                combined_result=combined_result,
            )
        )

        # Call the APIBackend to generate the response for hypothesis feedback
        response = APIBackend().build_messages_and_create_chat_completion(
            user_prompt=usr_prompt,
            system_prompt=sys_prompt,
            json_mode=True,
            json_target_type=Dict[str, str | bool | int],
        )

        # Parse the JSON response to extract the feedback
        response_json = json.loads(response)

        # Extract fields from JSON response
        observations = response_json.get("Observations", "No observations provided")
        hypothesis_evaluation = response_json.get("Feedback for Hypothesis", "No feedback provided")
        new_hypothesis = response_json.get("New Hypothesis", "No new hypothesis provided")
        reason = response_json.get("Reasoning", "No reasoning provided")
        decision = convert2bool(response_json.get("Replace Best Result", "no"))

        return HypothesisFeedback(
            observations=observations,
            hypothesis_evaluation=hypothesis_evaluation,
            new_hypothesis=new_hypothesis,
            reason=reason,
            decision=decision,
        )


class QlibModelExperiment2Feedback(Experiment2Feedback):
    """Generated feedbacks on the hypothesis from **Executed** Implementations of different tasks & their comparisons with previous performances"""

    def generate_feedback(self, exp: Experiment, trace: Trace) -> HypothesisFeedback:
        """
        The `ti` should be executed and the results should be included, as well as the comparison between previous results (done by LLM).
        For example: `mlflow` of Qlib will be included.
        """
        hypothesis = exp.hypothesis
        logger.info("Generating feedback...")
        # Define the system prompt for hypothesis feedback
        system_prompt = feedback_prompts["model_feedback_generation"]["system"]

        # Define the user prompt for hypothesis feedback
        context = trace.scen
        SOTA_hypothesis, SOTA_experiment = trace.get_sota_hypothesis_and_experiment()

        user_prompt = (
            Environment(undefined=StrictUndefined)
            .from_string(feedback_prompts["model_feedback_generation"]["user"])
            .render(
                context=context,
                last_hypothesis=SOTA_hypothesis,
                last_task=SOTA_experiment.sub_tasks[0].get_task_information() if SOTA_hypothesis else None,
                last_code=SOTA_experiment.sub_workspace_list[0].file_dict.get("model.py") if SOTA_hypothesis else None,
                last_result=SOTA_experiment.result if SOTA_hypothesis else None,
                hypothesis=hypothesis,
                exp=exp,
            )
        )

        # Call the APIBackend to generate the response for hypothesis feedback
        response_hypothesis = APIBackend().build_messages_and_create_chat_completion(
            user_prompt=user_prompt,
            system_prompt=system_prompt,
            json_mode=True,
            json_target_type=Dict[str, str | bool | int],
        )

        # Parse the JSON response to extract the feedback
        response_json_hypothesis = json.loads(response_hypothesis)
        return HypothesisFeedback(
            observations=response_json_hypothesis.get("Observations", "No observations provided"),
            hypothesis_evaluation=response_json_hypothesis.get("Feedback for Hypothesis", "No feedback provided"),
            new_hypothesis=response_json_hypothesis.get("New Hypothesis", "No new hypothesis provided"),
            reason=response_json_hypothesis.get("Reasoning", "No reasoning provided"),
            decision=convert2bool(response_json_hypothesis.get("Decision", "false")),
        )