feat: Dynamic scenario based on task (#392)

* OAI api

* more comments

* llm utils

* RD-Framework for dynamic scen

* remove useless code

* auto lint

* fix dynamic scennario

* Add parameter

* auto lint

* remove abstractmethod

* fix lint

rdagent/components/coder/factor_coder/CoSTEER/evaluators.py

@@ -89,7 +89,11 @@ class FactorCodeEvaluator(FactorEvaluator):
         system_prompt = (
             Environment(undefined=StrictUndefined)
             .from_string(evaluate_prompts["evaluator_code_feedback_v1_system"])
-            .render(scenario=self.scen.get_scenario_all_desc() if self.scen is not None else "No scenario description.")
+            .render(
+                scenario=self.scen.get_scenario_all_desc(target_task)
+                if self.scen is not None
+                else "No scenario description."
+            )
         )
 
         execution_feedback_to_render = execution_feedback
@@ -167,7 +171,11 @@ class FactorOutputFormatEvaluator(FactorEvaluator):
             .from_string(
                 evaluate_prompts["evaluator_output_format_system"],
             )
-            .render(scenario=self.scen.get_scenario_all_desc() if self.scen is not None else "No scenario description.")
+            .render(
+                scenario=self.scen.get_scenario_all_desc(implementation.target_task)
+                if self.scen is not None
+                else "No scenario description."
+            )
         )
 
         # TODO: with retry_context(retry_n=3, except_list=[KeyError]):
@@ -473,7 +481,11 @@ class FactorFinalDecisionEvaluator(Evaluator):
         system_prompt = (
             Environment(undefined=StrictUndefined)
             .from_string(evaluate_prompts["evaluator_final_decision_v1_system"])
-            .render(scenario=self.scen.get_scenario_all_desc() if self.scen is not None else "No scenario description.")
+            .render(
+                scenario=self.scen.get_scenario_all_desc(target_task)
+                if self.scen is not None
+                else "No scenario description."
+            )
         )
         execution_feedback_to_render = execution_feedback
 

rdagent/components/coder/factor_coder/CoSTEER/evolving_strategy.py

@@ -133,7 +133,7 @@ class FactorEvolvingStrategy(MultiProcessEvolvingStrategy):
                     implement_prompts["evolving_strategy_factor_implementation_v1_system"],
                 )
                 .render(
-                    scenario=self.scen.get_scenario_all_desc(),
+                    scenario=self.scen.get_scenario_all_desc(target_task),
                     queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
                 )
             )
@@ -226,7 +226,7 @@ class FactorEvolvingStrategyWithGraph(MultiProcessEvolvingStrategy):
                     implement_prompts["evolving_strategy_factor_implementation_v1_system"],
                 )
                 .render(
-                    scenario=self.scen.get_scenario_all_desc(),
+                    scenario=self.scen.get_scenario_all_desc(target_task),
                     queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
                 )
             )
@@ -250,7 +250,7 @@ class FactorEvolvingStrategyWithGraph(MultiProcessEvolvingStrategy):
                         Environment(undefined=StrictUndefined)
                         .from_string(implement_prompts["evolving_strategy_error_summary_v2_system"])
                         .render(
-                            scenario=self.scen.get_scenario_all_desc(),
+                            scenario=self.scen.get_scenario_all_desc(target_task),
                             factor_information_str=target_factor_task_information,
                             code_and_feedback=queried_former_failed_knowledge_to_render[
                                 -1

rdagent/components/coder/model_coder/CoSTEER/evaluators.py

@@ -89,7 +89,11 @@ class ModelCodeEvaluator(Evaluator):
         system_prompt = (
             Environment(undefined=StrictUndefined)
             .from_string(evaluate_prompts["evaluator_code_feedback"]["system"])
-            .render(scenario=self.scen.get_scenario_all_desc() if self.scen is not None else "No scenario description.")
+            .render(
+                scenario=self.scen.get_scenario_all_desc(target_task)
+                if self.scen is not None
+                else "No scenario description."
+            )
         )
 
         execution_feedback_to_render = model_execution_feedback
@@ -145,7 +149,11 @@ class ModelFinalEvaluator(Evaluator):
         system_prompt = (
             Environment(undefined=StrictUndefined)
             .from_string(evaluate_prompts["evaluator_final_feedback"]["system"])
-            .render(scenario=self.scen.get_scenario_all_desc() if self.scen is not None else "No scenario description.")
+            .render(
+                scenario=self.scen.get_scenario_all_desc(target_task)
+                if self.scen is not None
+                else "No scenario description."
+            )
         )
 
         execution_feedback_to_render = model_execution_feedback

rdagent/core/scenario.py

@@ -1,5 +1,7 @@
 from abc import ABC, abstractmethod
 
+from rdagent.core.experiment import Task
+
 
 class Scenario(ABC):
     @property
@@ -7,10 +9,21 @@ class Scenario(ABC):
     def background(self) -> str:
         """Background information"""
 
+    # TODO: We have to change all the sub classes to override get_source_data_desc instead of `source_data`
+    def get_source_data_desc(self, task: Task | None = None) -> str:  # noqa: ARG002
+        """
+        Source data description
+
+        The choice of data may vary based on the specific task at hand.
+        """
+        return ""
+
     @property
-    @abstractmethod
     def source_data(self) -> str:
-        """Source data description"""
+        """
+        A convenient shortcut for describing source data
+        """
+        return self.get_source_data_desc()
 
     @property
     @abstractmethod
@@ -33,8 +46,12 @@ class Scenario(ABC):
         """Rich style description to present"""
 
     @abstractmethod
-    def get_scenario_all_desc(self) -> str:
-        """Combine all the description together"""
+    def get_scenario_all_desc(self, task: Task | None = None) -> str:
+        """
+        Combine all descriptions together
+
+        The scenario description varies based on the task being performed.
+        """
 
     @property
     def experiment_setting(self) -> str | None:

rdagent/oai/llm_utils.py

@@ -235,6 +235,15 @@ class ChatSession:
 
 
 class APIBackend:
+    """
+    This is a unified interface for different backends.
+
+    (xiao) thinks integerate all kinds of API in a single class is not a good design.
+    So we should split them into different classes in `oai/backends/` in the future.
+    """
+
+    # FIXME: (xiao) We should avoid using self.xxxx.
+    # Instead, we can use self.cfg directly. If it's difficult to support different backend settings, we can split them into multiple BaseSettings.
     def __init__(  # noqa: C901, PLR0912, PLR0915
         self,
         *,

rdagent/scenarios/data_mining/experiment/model_experiment.py

@@ -5,6 +5,7 @@ from rdagent.components.coder.model_coder.model import (
     ModelFBWorkspace,
     ModelTask,
 )
+from rdagent.core.experiment import Task
 from rdagent.core.prompts import Prompts
 from rdagent.core.scenario import Scenario
 from rdagent.scenarios.data_mining.experiment.workspace import DMFBWorkspace
@@ -62,7 +63,7 @@ The demo showcases the iterative process of hypothesis generation, knowledge con
  
 To demonstrate the dynamic evolution of models through the R&D loop, emphasizing how each iteration enhances the model performance and reliability. The performane is measured by the AUROC score (Area Under the Receiver Operating Characteristic), which is a commonly used metric for binary classification.   """
 
-    def get_scenario_all_desc(self) -> str:
+    def get_scenario_all_desc(self, task: Task | None = None) -> str:
         return f"""Background of the scenario:
 {self.background}
 The interface you should follow to write the runnable code:

rdagent/scenarios/general_model/scenario.py

@@ -1,6 +1,7 @@
 from copy import deepcopy
 from pathlib import Path
 
+from rdagent.core.experiment import Task
 from rdagent.core.prompts import Prompts
 from rdagent.core.scenario import Scenario
 
@@ -40,7 +41,7 @@ class GeneralModelScenario(Scenario):
     def rich_style_description(self) -> str:
         return self._rich_style_description
 
-    def get_scenario_all_desc(self) -> str:
+    def get_scenario_all_desc(self, task: Task | None = None) -> str:
         return f"""Background of the scenario:
 {self.background}
 The interface you should follow to write the runnable code:

rdagent/scenarios/kaggle/experiment/scenario.py

@@ -8,6 +8,7 @@ import pandas as pd
 from jinja2 import Environment, StrictUndefined
 
 from rdagent.app.kaggle.conf import KAGGLE_IMPLEMENT_SETTING
+from rdagent.core.experiment import Task
 from rdagent.core.prompts import Prompts
 from rdagent.core.scenario import Scenario
 from rdagent.oai.llm_utils import APIBackend
@@ -203,7 +204,7 @@ The model code should follow the simulator:
 This is the Kaggle scenario for the competition: {self.competition}
 """
 
-    def get_scenario_all_desc(self) -> str:
+    def get_scenario_all_desc(self, task: Task | None = None) -> str:
         return f"""Background of the scenario:
 {self._background}
 The source dataset you can use to generate the features:

rdagent/scenarios/qlib/experiment/factor_experiment.py

@@ -6,6 +6,7 @@ from rdagent.components.coder.factor_coder.factor import (
     FactorFBWorkspace,
     FactorTask,
 )
+from rdagent.core.experiment import Task
 from rdagent.core.prompts import Prompts
 from rdagent.core.scenario import Scenario
 from rdagent.scenarios.qlib.experiment.utils import get_data_folder_intro
@@ -36,8 +37,7 @@ class QlibFactorScenario(Scenario):
     def background(self) -> str:
         return self._background
 
-    @property
-    def source_data(self) -> str:
+    def get_source_data_desc(self, task: Task | None = None) -> str:
         return self._source_data
 
     @property
@@ -60,11 +60,12 @@ class QlibFactorScenario(Scenario):
     def experiment_setting(self) -> str:
         return self._experiment_setting
 
-    def get_scenario_all_desc(self) -> str:
+    def get_scenario_all_desc(self, task: Task | None = None) -> str:
+        """A static scenario describer"""
         return f"""Background of the scenario:
 {self.background}
 The source data you can use:
-{self.source_data}
+{self.get_source_data_desc(task)}
 The interface you should follow to write the runnable code:
 {self.interface}
 The output of your code should be in the format:

rdagent/scenarios/qlib/experiment/model_experiment.py

@@ -6,6 +6,7 @@ from rdagent.components.coder.model_coder.model import (
     ModelFBWorkspace,
     ModelTask,
 )
+from rdagent.core.experiment import Task
 from rdagent.core.prompts import Prompts
 from rdagent.core.scenario import Scenario
 from rdagent.scenarios.qlib.experiment.workspace import QlibFBWorkspace
@@ -57,7 +58,7 @@ class QlibModelScenario(Scenario):
     def experiment_setting(self) -> str:
         return self._experiment_setting
 
-    def get_scenario_all_desc(self) -> str:
+    def get_scenario_all_desc(self, task: Task | None = None) -> str:
         return f"""Background of the scenario:
 {self.background}
 The interface you should follow to write the runnable code:

rdagent/scenarios/qlib/experiment/utils.py

@@ -1,3 +1,4 @@
+import re
 import shutil
 from pathlib import Path
 
@@ -49,15 +50,31 @@ def generate_data_folder_from_qlib():
     )
 
 
-def get_data_folder_intro():
-    """Directly get the info of the data folder.
+def get_data_folder_intro(fname_reg: str = ".*", flags=0) -> str:
+    """
+    Directly get the info of the data folder.
     It is for preparing prompting message.
+
+    Parameters
+    ----------
+    fname_reg : str
+        a regular expression to filter the file name.
+
+    flags: str
+        flags for re.match
+
+    Returns
+    -------
+        str
+            The description of the data folder.
     """
 
     if (
         not Path(FACTOR_IMPLEMENT_SETTINGS.data_folder).exists()
         or not Path(FACTOR_IMPLEMENT_SETTINGS.data_folder_debug).exists()
     ):
+        # FIXME: (xiao) I think this is writing in a hard-coded way.
+        # get data folder intro does not imply that we are generating the data folder.
         generate_data_folder_from_qlib()
 
     JJ_TPL = Environment(undefined=StrictUndefined).from_string(
@@ -70,29 +87,30 @@ def get_data_folder_intro():
     )
     content_l = []
     for p in Path(FACTOR_IMPLEMENT_SETTINGS.data_folder_debug).iterdir():
-        if p.name.endswith(".h5"):
-            df = pd.read_hdf(p)
-            # get  df.head() as string with full width
-            pd.set_option("display.max_columns", None)  # or 1000
-            pd.set_option("display.max_rows", None)  # or 1000
-            pd.set_option("display.max_colwidth", None)  # or 199
-            rendered = JJ_TPL.render(
-                file_name=p.name,
-                type_desc="generated by `pd.read_hdf(filename).head()`",
-                content=df.head().to_string(),
-            )
-            content_l.append(rendered)
-        elif p.name.endswith(".md"):
-            with open(p) as f:
-                content = f.read()
+        if re.match(fname_reg, p.name, flags) is not None:
+            if p.name.endswith(".h5"):
+                df = pd.read_hdf(p)
+                # get  df.head() as string with full width
+                pd.set_option("display.max_columns", None)  # or 1000
+                pd.set_option("display.max_rows", None)  # or 1000
+                pd.set_option("display.max_colwidth", None)  # or 199
                 rendered = JJ_TPL.render(
                     file_name=p.name,
-                    type_desc="markdown",
-                    content=content,
+                    type_desc="generated by `pd.read_hdf(filename).head()`",
+                    content=df.head().to_string(),
                 )
                 content_l.append(rendered)
-        else:
-            raise NotImplementedError(
-                f"file type {p.name} is not supported. Please implement its description function.",
-            )
-    return "\n ----------------- file splitter -------------\n".join(content_l)
+            elif p.name.endswith(".md"):
+                with open(p) as f:
+                    content = f.read()
+                    rendered = JJ_TPL.render(
+                        file_name=p.name,
+                        type_desc="markdown",
+                        content=content,
+                    )
+                    content_l.append(rendered)
+            else:
+                raise NotImplementedError(
+                    f"file type {p.name} is not supported. Please implement its description function.",
+                )
+    return "\n----------------- file splitter -------------\n".join(content_l)

rdagent/utils/agent/tpl.py

@@ -18,8 +18,11 @@ PROJ_PATH = DIRNAME.parent.parent
 
 
 # class T(SingletonBaseClass): TODO: singleton does not support args now.
-class T:
-    """Use the simplest way to (C)reate a Template and (r)ender it!!"""
+class RDAT:
+    """
+    RD-Agent's Template
+    Use the simplest way to (C)reate a Template and (r)ender it!!
+    """
 
     def __init__(self, uri: str):
         """
@@ -61,3 +64,6 @@ class T:
         Render the template with the given context.
         """
         return Environment(undefined=StrictUndefined).from_string(self.template).render(**context)
+
+
+T = RDAT  # shortcuts