from langchain import LLMChain
from langchain.agents import (AgentExecutor, Tool, ZeroShotAgent)
from langchain_experimental.tools import PythonREPLTool

# Define a description to suggest how to determine the choice of tool
description = (
    "Useful when you require to answer analytical questions about customers. "
    "Use this more than the Python REPL tool if the question is about customer analytics,"
    "like 'How many customers are there?' or 'count the number of transactions by age group'. "
    "Try not to use clause in the SQL."
)

# Create a Tool object for customer data with the previously defined agent executor 'create_sql_agent' and description
customer_data_tool = Tool(
    name="Customer",
    func=agent_executor.run,
    description=description,
)

# Create the whole list of tools
tools = [PythonREPLTool()]
tools.append(customer_data_tool)

# Define the prefix and suffix for the prompt
prefix = "Below are tools that you can access:"
suffix = (
    "Pass the relevant part of the request directly to the Customer tool.nn"
    "Request: {input}n"
    "{agent_scratchpad}"
)

# Create the prompt using ZeroShotAgent
# Use agent_scratchpad to store the actions previously used, guiding the subsequent responses.
agent_prompt = ZeroShotAgent.create_prompt(
    tools, prefix=prefix, suffix=suffix, input_variables=["input", "agent_scratchpad"]
)

# Create an instance of ZeroShotAgent with the LLMChain and the allowed tool names
zero_shot_agent = ZeroShotAgent(
    llm_chain=LLMChain(llm=llm, prompt=agent_prompt),
    allowed_tools=[tool.name for tool in tools]
)

# Create an AgentExecutor which enables verbose mode and handling parsing errors
agent_executor = AgentExecutor.from_agent_and_tools(
    agent=zero_shot_agent, tools=tools, verbose=True, handle_parsing_errors=True
)

# Define user input
user_inquiry = "Use a grouped bar graph to visualize the result of the following inquiry: " 
          "What are the relationships between product category, average total amount, and gender?"

# Run the agent to generate a response
agent_executor.run(user_inquiry)