feat: Added support for loading and storing RAG in Kaggle scenarios. (#…

…269)

* init a scenario for kaggle feature engineering

* Added support for loading and storing RAG in Kaggle scenarios.

* fix a ci bug

* Add RAG after each experiment's feedback.

* add a promt

* fix a bug

* fix a bug

* add a readme

* refine the code in knowledge loading

rdagent/app/kaggle/conf.py

@@ -10,7 +10,7 @@ class Config:
         env_prefix = "KG_"
         """Use `KG_` as prefix for environment variables"""
         protected_namespaces = ()
-        """Add 'model_' to the protected namespaces"""
+        """Do not allow overriding of these namespaces"""
 
     # 1) overriding the default
     scen: str = "rdagent.scenarios.kaggle.experiment.scenario.KGScenario"
@@ -42,5 +42,7 @@ class Config:
 
     competition: str = ""
 
+    rag_path: str = "git_ignore_folder/rag"
+
 
 KAGGLE_IMPLEMENT_SETTING = KaggleBasePropSetting()

rdagent/app/kaggle/loop.py

@@ -17,6 +17,9 @@
 from rdagent.core.scenario import Scenario
 from rdagent.core.utils import import_class
 from rdagent.log import rdagent_logger as logger
+from rdagent.scenarios.kaggle.knowledge_management.vector_base import (
+    KaggleExperienceBase,
+)
 from rdagent.scenarios.kaggle.proposal.proposal import (
     KG_ACTION_FEATURE_ENGINEERING,
     KG_ACTION_FEATURE_PROCESSING,
@@ -29,6 +32,10 @@ def __init__(self, PROP_SETTING: BasePropSetting):
             scen: Scenario = import_class(PROP_SETTING.scen)(PROP_SETTING.competition)
             logger.log_object(scen, tag="scenario")
 
+            self.vector_base = KaggleExperienceBase()
+            if KAGGLE_IMPLEMENT_SETTING.rag_path:
+                self.vector_base.load(KAGGLE_IMPLEMENT_SETTING.rag_path)
+
             self.hypothesis_gen: HypothesisGen = import_class(PROP_SETTING.hypothesis_gen)(scen)
             logger.log_object(self.hypothesis_gen, tag="hypothesis generator")
 

rdagent/scenarios/kaggle/developer/feedback.py

@@ -14,9 +14,12 @@
 )
 from rdagent.log import rdagent_logger as logger
 from rdagent.oai.llm_utils import APIBackend
+from rdagent.scenarios.kaggle.knowledge_management.extract_knowledge import (
+    extract_knowledge_from_feedback,
+)
 from rdagent.utils import convert2bool
 
-feedback_prompts = Prompts(file_path=Path(__file__).parent.parent / "prompts.yaml")
+prompt_dict = Prompts(file_path=Path(__file__).parent.parent / "prompts.yaml")
 DIRNAME = Path(__file__).absolute().resolve().parent
 
 
@@ -84,7 +87,7 @@ def generate_feedback(self, exp: Experiment, hypothesis: Hypothesis, trace: Trac
         # Generate the system prompt
         sys_prompt = (
             Environment(undefined=StrictUndefined)
-            .from_string(feedback_prompts["factor_feedback_generation"]["system"])
+            .from_string(prompt_dict["factor_feedback_generation"]["system"])
             .render(scenario=self.scen.get_scenario_all_desc())
         )
 
@@ -97,7 +100,7 @@ def generate_feedback(self, exp: Experiment, hypothesis: Hypothesis, trace: Trac
         # Generate the user prompt
         usr_prompt = (
             Environment(undefined=StrictUndefined)
-            .from_string(feedback_prompts[prompt_key]["user"])
+            .from_string(prompt_dict[prompt_key]["user"])
             .render(
                 hypothesis_text=hypothesis_text,
                 task_details=tasks_factors,
@@ -122,6 +125,17 @@ def generate_feedback(self, exp: Experiment, hypothesis: Hypothesis, trace: Trac
         reason = response_json.get("Reasoning", "No reasoning provided")
         decision = convert2bool(response_json.get("Replace Best Result", "no"))
 
+        experiment_feedback = {
+            "hypothesis_text": hypothesis_text,
+            "current_result": current_result,
+            "tasks_factors": tasks_factors,
+            "observations": observations,
+            "hypothesis_evaluation": hypothesis_evaluation,
+            "reason": reason,
+        }
+
+        self.scen.vector_base.add_experience_to_vector_base(experiment_feedback)
+
         return HypothesisFeedback(
             observations=observations,
             hypothesis_evaluation=hypothesis_evaluation,

rdagent/scenarios/kaggle/experiment/scenario.py

@@ -4,12 +4,16 @@
 import pandas as pd
 from jinja2 import Environment, StrictUndefined
 
+from rdagent.app.kaggle.conf import KAGGLE_IMPLEMENT_SETTING
 from rdagent.components.coder.factor_coder.config import FACTOR_IMPLEMENT_SETTINGS
 from rdagent.core.prompts import Prompts
 from rdagent.core.scenario import Scenario
 from rdagent.oai.llm_utils import APIBackend
 from rdagent.scenarios.kaggle.experiment.kaggle_experiment import KGFactorExperiment
 from rdagent.scenarios.kaggle.kaggle_crawler import crawl_descriptions
+from rdagent.scenarios.kaggle.knowledge_management.vector_base import (
+    KaggleExperienceBase,
+)
 
 prompt_dict = Prompts(file_path=Path(__file__).parent / "prompts.yaml")
 
@@ -32,6 +36,11 @@ def __init__(self, competition: str) -> None:
 
         self._background = self.background
 
+        # all competitions are based on the same vector base
+        self.vector_base = KaggleExperienceBase()
+        if KAGGLE_IMPLEMENT_SETTING.rag_path:
+            self.vector_base.load(KAGGLE_IMPLEMENT_SETTING.rag_path)
+
     def _analysis_competition_description(self):
         sys_prompt = (
             Environment(undefined=StrictUndefined)

rdagent/scenarios/kaggle/knowledge_management/README.md

@@ -0,0 +1,8 @@
+## Usage
+
+This folder implements a knowledge base using RAG based on Kaggle competitions. 
+It allows you to store Kaggle competition experiences into the knowledge base, as well as store experimental experiences from RD-Agent.
+
+1. First, generate a knowledge base (in JSON format) by running the `main` function in `extract_knowledge.py`.
+2. Then, create a vector base in `vector_base.py` and save it.
+3. Finally, add the field `KG_RAG_PATH="xxx.pkl"` (the path to the saved vector base) in your `.env` file.

rdagent/scenarios/kaggle/knowledge_management/extract_experience_from_docs.py → rdagent/scenarios/kaggle/knowledge_management/extract_knowledge.py

@@ -10,7 +10,7 @@
 prompt_dict = Prompts(file_path=Path(__file__).parent / "prompts.yaml")
 
 
-def process_with_gpt(content: str):
+def extract_knowledge_from_high_score_answers(content: str):
     sys_prompt = (
         Environment(undefined=StrictUndefined)
         .from_string(prompt_dict["extract_kaggle_knowledge_prompts"]["system"])
@@ -37,6 +37,36 @@ def process_with_gpt(content: str):
     return response_json_analysis
 
 
+def extract_knowledge_from_feedback(feedback_response: dict) -> dict:
+    """
+    Extracts knowledge from LLM-generated feedback and structures it.
+    """
+    sys_prompt = (
+        Environment(undefined=StrictUndefined)
+        .from_string(prompt_dict["extract_kaggle_knowledge_from_feedback_prompts"]["system"])
+        .render()
+    )
+
+    user_prompt = (
+        Environment(undefined=StrictUndefined)
+        .from_string(prompt_dict["extract_kaggle_knowledge_from_feedback_prompts"]["user"])
+        .render(experiment_strategy=feedback_response)
+    )
+
+    response_analysis = APIBackend().build_messages_and_create_chat_completion(
+        user_prompt=user_prompt,
+        system_prompt=sys_prompt,
+        json_mode=True,
+    )
+
+    try:
+        response_json_analysis = json.loads(response_analysis)
+    except json.JSONDecodeError:
+        response_json_analysis = {"error": "Failed to parse LLM's response as JSON"}
+
+    return response_json_analysis
+
+
 def process_all_case_files(directory_path: str):
     output_file = Path(directory_path) / "kaggle_experience_results.json"
     json_output = []
@@ -46,8 +76,8 @@ def process_all_case_files(directory_path: str):
 
             with open(file_path, "r", encoding="utf-8") as file:
                 content = file.read()
-                gpt_response = process_with_gpt(content)
-                json_output.append(gpt_response)
+                knowladge = extract_knowledge_from_high_score_answers(content)
+                json_output.append(knowladge)
 
     with open(output_file, "w", encoding="utf-8") as json_file:
         json.dump(json_output, json_file, ensure_ascii=False)

rdagent/scenarios/kaggle/knowledge_management/prompts.yaml

@@ -16,4 +16,24 @@ extract_kaggle_knowledge_prompts:
     }
   
   user: |-
-    High-ranking Kaggle notebooks or competition strategies: {{ file_content }}
+    High-ranking Kaggle notebooks or competition strategies: {{ file_content }}
+
+extract_kaggle_knowledge_from_feedback_prompts:
+  system: |-
+    You are a Kaggle competition expert with extensive experience in analyzing Kaggle notebooks and competition strategies. 
+    Your task is to summarize or infer key information such as the competition name, task type, and specific techniques employed in the notebook or strategy.
+    For each provided content, you are expected to extract valuable insights and organize the analysis in the structured format outlined below.
+    
+    Please provide the analysis in the following JSON format:
+    {
+      "content": "all provided content",
+      "title": "extracted title, if available",
+      "competition_name": "extracted competition name",
+      "task_category": "extracted task type, e.g., Classification, Regression",
+      "field": "field of focus, e.g., Feature Engineering, Modeling",
+      "ranking": "extracted ranking, if available",
+      "score": "extracted score or metric, if available"
+    }
+  
+  user: |-
+    Experiment strategy: {{ experiment_strategy }}

rdagent/scenarios/kaggle/knowledge_management/vector_base.py

@@ -12,6 +12,9 @@
 )
 from rdagent.log import rdagent_logger as logger
 from rdagent.oai.llm_utils import APIBackend
+from rdagent.scenarios.kaggle.knowledge_management.extract_knowledge import (
+    extract_knowledge_from_feedback,
+)
 
 
 class KGKnowledgeMetaData(KnowledgeMetaData):
@@ -178,10 +181,32 @@ def load_kaggle_experience(self, kaggle_experience_path: Union[str, Path]):
             logger.error(f"Kaggle experience data not found at {kaggle_experience_path}")
             self.kaggle_experience_data = []
 
-    def add_experience_to_vector_base(self):
+    def add_experience_to_vector_base(self, experiment_feedback=None):
         """
-        Process the Kaggle experience data and add relevant information to the vector base
+        Process Kaggle experience data or experiment feedback and add relevant information to the vector base.
+
+        Args:
+            experiment_feedback (dict, optional): A dictionary containing experiment feedback.
+                                                If provided, this feedback will be processed and added to the vector base.
         """
+        # If experiment feedback is provided, extract relevant knowledge and add it to the vector base
+        if experiment_feedback:
+            extracted_knowledge = extract_knowledge_from_feedback(experiment_feedback)
+
+            document = KGKnowledgeMetaData(
+                content=experiment_feedback.get("hypothesis_text", ""),
+                label="Experiment Feedback",
+                competition_name="Experiment Result",
+                task_category=experiment_feedback.get("tasks_factors", "General Task"),
+                field="Research Feedback",
+                ranking=None,
+                score=experiment_feedback.get("current_result", None),
+            )
+            document.create_embedding()
+            self.add(document)
+            return
+
+        # Process Kaggle experience data
         for experience in self.kaggle_experience_data:
             content = experience.get("content", "")
             label = experience.get("title", "Kaggle Experience")
@@ -238,6 +263,8 @@ def search_experience(self, query: str, topk_k: int = 5, similarity_threshold: f
 
     kaggle_base.add_experience_to_vector_base()
 
+    kaggle_base.save("git_ignore_folder/experience/tabular_cases/kaggle_vector_base.pkl")
+
     print(f"There are {kaggle_base.shape()[0]} records in the vector base.")
 
     search_results, similarities = kaggle_base.search_experience(query="image classification", topk_k=3)

rdagent/scenarios/kaggle/proposal/proposal.py

@@ -4,8 +4,10 @@
 
 from jinja2 import Environment, StrictUndefined
 
+from rdagent.app.kaggle.conf import KAGGLE_IMPLEMENT_SETTING
 from rdagent.components.coder.factor_coder.factor import FactorTask
 from rdagent.components.coder.model_coder.model import ModelExperiment, ModelTask
+from rdagent.components.knowledge_management.vector_base import VectorBase
 from rdagent.components.proposal.model_proposal import (
     ModelHypothesis,
     ModelHypothesis2Experiment,
@@ -17,6 +19,9 @@
     KGFactorExperiment,
     KGModelExperiment,
 )
+from rdagent.scenarios.kaggle.knowledge_management.vector_base import (
+    KaggleExperienceBase,
+)
 
 prompt_dict = Prompts(file_path=Path(__file__).parent.parent / "prompts.yaml")
 
@@ -68,22 +73,27 @@ class KGHypothesisGen(ModelHypothesisGen):
 
     .. code-block:: python
 
-        class XXXDMModelHypothesisGen(DMModelHypothesisGen):
+        class KGHypothesisGen(ModelHypothesisGen):
             prompts: Prompts = a_specifc_prompt_dict
     """
 
-    def __init__(self, scen: Scenario) -> Tuple[dict, bool]:
+    def __init__(self, scen: Scenario, knowledge: VectorBase = None) -> Tuple[dict, bool]:
         super().__init__(scen)
+        self.scen.vector_base.save(KAGGLE_IMPLEMENT_SETTING.rag_path)
 
     def prepare_context(self, trace: Trace) -> Tuple[dict, bool]:
         hypothesis_feedback = (
             Environment(undefined=StrictUndefined)
             .from_string(prompt_dict["hypothesis_and_feedback"])
             .render(trace=trace)
         )
+
+        rag_results, _ = self.scen.vector_base.search_experience(hypothesis_feedback, topk_k=5)
+        rag_content = "\n".join([doc.content for doc in rag_results])
+
         context_dict = {
             "hypothesis_and_feedback": hypothesis_feedback,
-            "RAG": None,
+            "RAG": rag_content,
             "hypothesis_output_format": prompt_dict["hypothesis_output_format"],
             "hypothesis_specification": None,
         }