Code cot

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

@@ -23,6 +23,7 @@
 from rdagent.core.prompts import Prompts
 from rdagent.core.utils import multiprocessing_wrapper
 from rdagent.oai.llm_utils import APIBackend
+from rdagent.utils.agent.tpl import T
 
 if TYPE_CHECKING:
     from rdagent.components.coder.factor_coder.CoSTEER.knowledge_management import (
@@ -127,34 +128,22 @@ def implement_one_factor(
 
             queried_former_failed_knowledge_to_render = queried_former_failed_knowledge
 
-            system_prompt = (
-                Environment(undefined=StrictUndefined)
-                .from_string(
-                    implement_prompts["evolving_strategy_factor_implementation_v1_system"],
-                )
-                .render(
+            system_prompt = T(".prompts:evolving_strategy_factor_implementation_v1_system").r(
                     scenario=self.scen.get_scenario_all_desc(),
                     queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
-                )
+                    enable_code_cot=FACTOR_IMPLEMENT_SETTINGS.enable_code_cot,
             )
             session = APIBackend(use_chat_cache=FACTOR_IMPLEMENT_SETTINGS.coder_use_cache).build_chat_session(
                 session_system_prompt=system_prompt,
             )
 
             queried_similar_successful_knowledge_to_render = queried_similar_successful_knowledge
             for _ in range(10):  # max attempt to reduce the length of user_prompt
-                user_prompt = (
-                    Environment(undefined=StrictUndefined)
-                    .from_string(
-                        implement_prompts["evolving_strategy_factor_implementation_v1_user"],
-                    )
-                    .render(
+                user_prompt = T(".prompts:evolving_strategy_factor_implementation_v1_user").r(
                         factor_information_str=factor_information_str,
                         queried_similar_successful_knowledge=queried_similar_successful_knowledge_to_render,
                         queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
-                    )
-                    .strip("\n")
-                )
+                    ).strip("\n")
                 if (
                     session.build_chat_completion_message_and_calculate_token(
                         user_prompt,
@@ -228,6 +217,7 @@ def implement_one_factor(
                 .render(
                     scenario=self.scen.get_scenario_all_desc(),
                     queried_former_failed_knowledge=queried_former_failed_knowledge_to_render,
+                    enable_code_cot=FACTOR_IMPLEMENT_SETTINGS.enable_code_cot,
                 )
             )
 

rdagent/components/coder/factor_coder/config.py

@@ -52,6 +52,9 @@ class Config:
     max_loop: int = 10
     """Maximum number of task implementation loops"""
 
+    enable_code_cot: bool = False
+    """Indicates whether to enable code cot"""
+
     knowledge_base_path: Union[str, None] = None
     """Path to the knowledge base"""
 

rdagent/components/coder/factor_coder/prompts.yaml

@@ -57,9 +57,46 @@ evolving_strategy_factor_implementation_v1_system: |-
   {{ queried_former_failed_knowledge[-1].feedback }}
   {% endif %}
 
+  {% if enable_code_cot %}
+  -------------- **The code standard** ----------------
+  You must write code with detailed comments to explain your thought process!!!
+  Even if other example code does not follow this, you should strictly adhere to this code standard.
+  If data processing is involved, include comments to describe the index and columns at each step.
+  Here is an example:
+  ```python
+  import pandas as pd
+  focused_fields = ["NET_PROFIT_INCL_MIN_INT_INC"]
+
+  df = pd.read_hdf("ASHAREINCOME.h5", key="data")  # type: pd.DataFrame; index with multi-level [datetime, instrument]; columns with single level that contains values like [REPORT_PERIOD, STATEMENT_TYPE,  ...,  NET_PROFIT_INCL_MIN_INT_INC]
+
+  # filter only part of the report statement
+  df = df[df["STATEMENT_TYPE"] == "408006000"]  # type: pd.DataFrame; index with multi-level [datetime, instrument]; columns with single level that contains values like [REPORT_PERIOD, STATEMENT_TYPE,  ...,  NET_PROFIT_INCL_MIN_INT_INC]
+  df = df[["REPORT_PERIOD"] + focused_fields]  # type: pd.DataFrame; index with multi-level [datetime, instrument]; columns with single level that contains values like [REPORT_PERIOD, NET_PROFIT_INCL_MIN_INT_INC]
+  df.columns.name = "field"  # type: pd.DataFrame; index with multi-level [datetime, instrument]; columns level [filed] that contains values like [REPORT_PERIOD, NET_PROFIT_INCL_MIN_INT_INC]
+  df = df.set_index("REPORT_PERIOD", append=True).unstack(level="REPORT_PERIOD")  # type: pd.DataFrame; index with multi-level [datetime, instrument]; columns mulit-level [filed, REPORT_PERIOD]
+  df = df.unstack("instrument")  # type: pd.DataFrame; index with level [datetime]; columns mulit-level [filed, REPORT_PERIOD,  instrument]
+  # fill the previous published value to later release date
+  df = df.sort_index().ffill(axis=0)  # type: pd.DataFrame; index with level [datetime]; columns mulit-level [filed, REPORT_PERIOD,  instrument]
+
+  pit_f_final = {}
+  for idx, row in df.iterrows():
+      pit_df = row.unstack("field").sort_index()  # ensure the report_period is ascending order  # type: pd.DataFrame; index with level [REPORT_PERIOD,  instrument]; columns mulit-level [filed]
+      # 1) collapse all the columns
+      pit_f = pit_df["NET_PROFIT_INCL_MIN_INT_INC"]  # type: pd.Series; index with level [REPORT_PERIOD,  instrument]
+      # 2)
+      pit_f_df = pit_f.unstack("instrument")   # type: pd.DataFrame; index with level [REPORT_PERIOD]; columns multi-level [instrument]
+      pit_f_ttm = pit_f_df.iloc[-4:, :].sum()  # this for TTM  # type: pd.Series; index with level [instrument]
+      pit_f_final[idx] = pit_f_ttm
+
+  pit_f_final = pd.DataFrame(pit_f_final).T  # type: pd.DataFrame; index with level [datetime]; columns multi-level [instrument]
+  pit_f_final = pit_f_final.stack()  # type: pd.Series; index with level [datetime, instrument]
+  pit_f_final.to_frame("net_profit_ttm").to_hdf('result.h5', key='data', mode='w')
+  ```
+  {% endif %}
+
   Please response the code in the following json format. Here is an example structure for the JSON output:
   {
-      "code": "The Python code as a string."
+      "code": "The Python code as a string that follow the code standard."
   }
 
 evolving_strategy_factor_implementation_v1_user: |-
@@ -86,6 +123,7 @@ evolving_strategy_factor_implementation_v1_user: |-
   {% endfor %}
   {% endif %}
 
+
 evolving_strategy_factor_implementation_v2_user: |-
   --------------Target factor information:---------------
   {{ factor_information_str }}

rdagent/utils/agent/tpl.py

@@ -17,6 +17,14 @@
 PROJ_PATH = DIRNAME.parent.parent
 
 
+# TODO: It could be better to add the Truncator class into  T class for reusing.
+# class Truncator:
+#     def __init__(self, limit: int):
+#         self.limit = limit
+#
+#     def render_with_trunc(self, render_func, context: dict):
+
+
 # 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!!"""