Introspective agents

bootstraprag/templates/llamaindex/agents_with_introspection/.env

@@ -0,0 +1 @@
+OPENAI_API_KEY='sk-proj-'

bootstraprag/templates/llamaindex/agents_with_introspection/agent_core.py

@@ -0,0 +1,92 @@
+from llama_index.core.tools.tool_spec.base import BaseToolSpec
+from llama_index.agent.introspective import ToolInteractiveReflectionAgentWorker, IntrospectiveAgentWorker
+from llama_index.core.agent import FunctionCallingAgentWorker  # Import OpenAIAgentWorker here
+from llama_index.agent.openai import OpenAIAgentWorker
+from llama_index.llms.openai import OpenAI
+from llama_index.core.llms import ChatMessage, MessageRole
+from datetime import datetime
+import yfinance as yf
+import pandas as pd
+from dotenv import load_dotenv, find_dotenv
+
+load_dotenv(find_dotenv())
+
+
+class StockDataRetrieverTool(BaseToolSpec):
+    """Tool for retrieving stock data"""
+
+    spec_functions = ['fetch_historical_prices']
+
+    def fetch_historical_prices(self, ticker: str, start_date: str,
+                                end_date: str = datetime.today().strftime('%Y-%m-%d')) -> pd.DataFrame:
+        """
+        Retrieves historical stock prices for a given ticker within a date range.
+
+        Args:
+            ticker (str): The stock ticker symbol.
+            start_date (str): Start date for data in 'YYYY-MM-DD' format.
+            end_date (str, optional): End date in 'YYYY-MM-DD' format. Defaults to today.
+
+        Returns:
+            pd.DataFrame: DataFrame with stock price data.
+        """
+        return yf.download(tickers=ticker, start=start_date, end=end_date, progress=True)
+
+
+class FinancialAgentBuilder:
+    """Builder for creating an introspective financial agent with reflection capabilities"""
+
+    def __init__(self, data_tool: BaseToolSpec):
+        self.data_tool = data_tool
+        self.data_tool_list = data_tool.to_tool_list()
+
+    def create_introspective_agent(self, verbose: bool = True, include_main_worker: bool = True):
+        critique_worker = self._build_critique_worker(verbose)
+        reflection_worker = self._build_reflection_worker(critique_worker, verbose)
+        main_worker = self._build_main_worker() if include_main_worker else None
+
+        introspective_worker = IntrospectiveAgentWorker.from_defaults(
+            reflective_agent_worker=reflection_worker,
+            main_agent_worker=main_worker,
+            verbose=verbose
+        )
+
+        initial_message = [
+            ChatMessage(
+                content="You are a financial assistant that helps gather historical prices "
+                        "and provides a summary of the closing price analysis.",
+                role=MessageRole.SYSTEM,
+            )
+        ]
+        return introspective_worker.as_agent(chat_history=initial_message, verbose=verbose)
+
+    def _build_critique_worker(self, verbose: bool):
+        return FunctionCallingAgentWorker.from_tools(
+            tools=self.data_tool_list,
+            llm=OpenAI(model="gpt-4o-mini"),
+            verbose=verbose
+        )
+
+    def _build_reflection_worker(self, critique_worker, verbose: bool):
+        def stop_critique_on_pass(critique_str: str):
+            return "[APPROVE]" in critique_str
+
+        return ToolInteractiveReflectionAgentWorker.from_defaults(
+            critique_agent_worker=critique_worker,
+            critique_template=(
+                "Please review the retrieval of historical prices with the specified time intervals. "
+                "Evaluate the price for accuracy, efficiency, and adherence to truth. "
+                "If everything is correct, write '[APPROVE]'; otherwise, write '[REJECT]'.\n\n{input_str}"
+            ),
+            stopping_callable=stop_critique_on_pass,
+            correction_llm=OpenAI(model='gpt-4o'),
+            verbose=verbose
+        )
+
+    def _build_main_worker(self):
+        return OpenAIAgentWorker.from_tools(
+            tools=self.data_tool_list,
+            llm=OpenAI("gpt-4-turbo"),
+            verbose=True
+        )
+

bootstraprag/templates/llamaindex/agents_with_introspection/api_server.py

@@ -1,23 +1,37 @@
 import litserve as lit
+from agent_core import StockDataRetrieverTool, FinancialAgentBuilder
+
+# Query example
+# query = """I have 10k dollars Now analyze APPL stock and MSFT stock to let me know where to invest
+#             this money and how many stock will I get it, Give me the last 3 months historical close prices of
+#             APPL and MSFT. Respond with a comparative summary on closing prices and recommended stock to invest."""
 
 
 class IntrospectionAgentAPI(lit.LitAPI):
     def __init__(self):
-        pass
+        # Instantiate tool and agent
+        self.stock_data_tool = StockDataRetrieverTool()
+        self.financial_agent_builder = None
 
     def setup(self, device):
-        pass
+        self.financial_agent_builder = FinancialAgentBuilder(data_tool=self.stock_data_tool)
 
     def decode_request(self, request, **kwargs):
-        pass
+        return request["query"]
 
     def predict(self, x, **kwargs):
-        pass
+        introspective_agent = self.financial_agent_builder.create_introspective_agent(verbose=True)
+        response = introspective_agent.chat(x)
+        return response
 
     def encode_response(self, output, **kwargs):
-        pass
+        return {"agent": output}
 
 
 if __name__ == "__main__":
     api = IntrospectionAgentAPI()
-    lit.LitServer(lit_api=api, spec=lit.OpenAISpec())
+    # pass this parameter "spec=lit.OpenAISpec()", if you want to use the API as
+    # http://localhost:8002/v1/chat/completion
+    # else http://localhost:8002/predict
+    server = lit.LitServer(api, accelerator="auto", spec=lit.OpenAISpec())
+    server.run(port=8002)

bootstraprag/templates/llamaindex/agents_with_introspection/main.py

@@ -0,0 +1,13 @@
+from agent_core import StockDataRetrieverTool, FinancialAgentBuilder
+
+# Instantiate tool and agent
+stock_data_tool = StockDataRetrieverTool()
+financial_agent_builder = FinancialAgentBuilder(data_tool=stock_data_tool)
+introspective_agent = financial_agent_builder.create_introspective_agent(verbose=True)
+
+# Query example
+query = """I have 10k dollars Now analyze APPL stock and MSFT stock to let me know where to invest 
+            this money and how many stock will I get it, Give me the last 3 months historical close prices of APPL and MSFT. 
+            Respond with a comparative summary on closing prices and recommended stock to invest."""
+response = introspective_agent.chat(query)
+print(response)

bootstraprag/templates/llamaindex/agents_with_introspection/readme.md

@@ -1 +1,20 @@
-### under progress
+### Introspective Financial Agents
+
+Introspective agents are artificial intelligence systems designed to evaluate and refine their own outputs through iterative reflection and correction. This self-assessment process enables them to improve performance over time. For instance, in the LlamaIndex framework, an introspective agent generates an initial response to a task and then engages in reflection and correction cycles to enhance the response until a satisfactory outcome is achieved.
+
+An introspective worker agent is a specific implementation of this concept. It operates by delegating tasks to two other agents: one generates the initial response, and the other performs reflection and correction on that response. This structured approach ensures systematic evaluation and improvement of outputs.
+
+### How to run ?
+- `pip install -r requirements.txt`
+- `python main.py` [standalone code]
+- `python api_server.py` [litserve apis]
+
+### How to access API?
+- URL: http://localhost:8002/predict
+- method: POST
+- payload: 
+```json
+{
+"query": "I have 10k dollars Now analyze APPL stock and MSFT stock to let me know where to invest this money and how many stock will I get it, Give me the last 3 months historical close prices of APPL and MSFT. Respond with a comparative summary on closing prices and recommended stock to invest."
+}
+```