app.py

import streamlit as st
import pandas as pd
import os
from crewai import Agent, Task, Crew
from langchain_groq import ChatGroq
import streamlit_ace as st_ace
import traceback
import contextlib
import io
from crewai_tools import FileReadTool
import matplotlib.pyplot as plt
import glob
from dotenv import load_dotenv

# load the .env file
load_dotenv()
# Set up Groq API key
groq_api_key = os.getenv("GROQ_API_KEY")


def main():
    # Set custom CSS for UI
    set_custom_css()

    # Initialize session state for edited code
    if 'edited_code' not in st.session_state:
        st.session_state['edited_code'] = ""
    
    # Initialize session state for whether the initial code is generated
    if 'code_generated' not in st.session_state:
        st.session_state['code_generated'] = False

    # Header with futuristic design
    st.markdown("""
        <div class="header">
            <h1>CrewAI Machine Learning Assistant</h1>
            <p>Your AI-powered partner for machine learning projects.</p>
        </div>
    """, unsafe_allow_html=True)

    # Sidebar for customization options
    st.sidebar.title('Customization')
    model = st.sidebar.selectbox(
        'Choose a model',
        ['llama3-8b-8192', "llama3-70b-8192"]
    )

    # Initialize LLM
    llm = initialize_llm(model)

    

    # User inputs
    user_question = st.text_area("Describe your ML problem:", key="user_question")
    uploaded_file = st.file_uploader("Upload a sample .csv of your data (optional)", key="uploaded_file")
    try:
        file_name = uploaded_file.name
    except: 
        file_name = "dataset.csv"

    # Initialize agents
    agents = initialize_agents(llm,file_name)
    # Process uploaded file
    if uploaded_file:
        try:
            df = pd.read_csv(uploaded_file)
            st.write("Data successfully uploaded:")
            st.dataframe(df.head())
            data_upload = True
        except Exception as e:
            st.error(f"Error reading the file: {e}")
            data_upload = False
    else:
        df = None
        data_upload = False

    # Process button
    if st.button('Process'):
        tasks = create_tasks("Process",user_question,file_name, data_upload, df, None, st.session_state['edited_code'], None, agents)
        with st.spinner('Processing...'):
            crew = Crew(
                agents=list(agents.values()),
                tasks=tasks,
                verbose=2
            )

            result = crew.kickoff()

            if result:  # Only call st_ace if code has a valid value
                code = result.strip("```")
                try:
                    filt_idx = code.index("```")
                    code = code[:filt_idx]
                except:
                    pass
                st.session_state['edited_code'] = code
                st.session_state['code_generated'] = True

        st.session_state['edited_code'] = st_ace.st_ace(
            value=st.session_state['edited_code'],
            language='python',
            theme='monokai',
            keybinding='vscode',
            min_lines=20,
            max_lines=50
        )

    if st.session_state['code_generated']:
        # Show options for modification, debugging, and running the code
        suggestion = st.text_area("Suggest modifications to the generated code (optional):", key="suggestion")
        if st.button('Modify'):
            if st.session_state['edited_code'] and suggestion:
                tasks = create_tasks("Modify",user_question,file_name, data_upload, df, suggestion, st.session_state['edited_code'], None, agents)
                with st.spinner('Modifying code...'):
                    crew = Crew(
                        agents=list(agents.values()),
                        tasks=tasks,
                        verbose=2
                    )

                    result = crew.kickoff()

                    if result:  # Only call st_ace if code has a valid value
                        code = result.strip("```")
                        try:
                            filter_idx = code.index("```")
                            code = code[:filter_idx]
                        except:
                            pass
                        st.session_state['edited_code'] = code

                st.write("Modified code:")
                st.session_state['edited_code']= st_ace.st_ace(
                    value=st.session_state['edited_code'],
                    language='python',
                    theme='monokai',
                    keybinding='vscode',
                    min_lines=20,
                    max_lines=50
                )

        debugger = st.text_area("Paste error message here for debugging (optional):", key="debugger")
        if st.button('Debug'):
            if st.session_state['edited_code'] and debugger:
                tasks = create_tasks("Debug",user_question,file_name, data_upload, df, None, st.session_state['edited_code'], debugger, agents)
                with st.spinner('Debugging code...'):
                    crew = Crew(
                        agents=list(agents.values()),
                        tasks=tasks,
                        verbose=2
                    )

                    result = crew.kickoff()

                    if result:  # Only call st_ace if code has a valid value
                        code = result.strip("```")
                        try:
                            filter_idx = code.index("```")
                            code = code[:filter_idx]
                        except:
                            pass
                        st.session_state['edited_code'] = code

                st.write("Debugged code:")
                st.session_state['edited_code'] = st_ace.st_ace(
                    value=st.session_state['edited_code'],
                    language='python',
                    theme='monokai',
                    keybinding='vscode',
                    min_lines=20,
                    max_lines=50
                )

        if st.button('Run'):
            output = io.StringIO()
            with contextlib.redirect_stdout(output):
                try:
                    globals().update({'dataset': df})
                    final_code = st.session_state["edited_code"]
                    
                    with st.expander("Final Code"):
                        st.code(final_code, language='python')

                    exec(final_code, globals())
                    result = output.getvalue()
                    success = True
                except Exception as e:
                    result = str(e)
                    success = False

            st.subheader('Output:')
            st.text(result)

            figs = [manager.canvas.figure for manager in plt._pylab_helpers.Gcf.get_all_fig_managers()]
            if figs:
                st.subheader('Generated Plots:')
                for fig in figs:
                    st.pyplot(fig)

            if success:
                st.success("Code executed successfully!")
            else:
                st.error("Code execution failed! Waiting for debugging input...")

            # Move the generated files section to the sidebar
            with st.sidebar:
                st.header('Output Files:')
                files = glob.glob(os.path.join("Output/", '*'))
                for file in files:
                    if os.path.isfile(file):
                        with open(file, 'rb') as f:
                            st.download_button(label=f'Download {os.path.basename(file)}', data=f, file_name=os.path.basename(file))



# Function to set custom CSS for futuristic UI
def set_custom_css():
    st.markdown("""
        <style>
            body {
                background: #0e0e0e;
                color: #e0e0e0;
                font-family: 'Roboto', sans-serif;
            }
            .header {
                background: linear-gradient(135deg, #6e3aff, #b839ff);
                padding: 10px;
                border-radius: 10px;
            }
            .header h1, .header p {
                color: white;
                text-align: center;
            }
            .stButton button {
                background-color: #b839ff;
                color: white;
                border-radius: 10px;
                font-size: 16px;
                padding: 10px 20px;
            }
            .stButton button:hover {
                background-color: #6e3aff;
                color: #e0e0e0;
            }
            .spinner {
                display: flex;
                justify-content: center;
                align-items: center;
            }
        </style>
    """, unsafe_allow_html=True)

# Function to initialize LLM
def initialize_llm(model):
    return ChatGroq(
        temperature=0,
        groq_api_key=groq_api_key,
        model_name=model
    )

# Function to initialize agents
def initialize_agents(llm,file_name):
    file_read_tool = FileReadTool()
    return {
        "Data_Reader_Agent": Agent(
            role='Data_Reader_Agent',
            goal="Read the uploaded dataset and provide it to other agents.",
            backstory="Responsible for reading the uploaded dataset.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
            tools=[file_read_tool]
        ),
        "Problem_Definition_Agent": Agent(
            role='Problem_Definition_Agent',
            goal="Clarify the machine learning problem the user wants to solve.",
            backstory="Expert in defining machine learning problems.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        "EDA_Agent": Agent(
            role='EDA_Agent',
            goal="Perform all possible Exploratory Data Analysis (EDA) on the data provided by the user.",
            backstory="Specializes in conducting comprehensive EDA to understand the data characteristics, distributions, and relationships.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        "Feature_Engineering_Agent": Agent(
            role='Feature_Engineering_Agent',
            goal="Perform feature engineering on the data based on the EDA results provided by the EDA agent.",
            backstory="Expert in deriving new features, transforming existing features, and preprocessing data to prepare it for modeling.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        "Model_Recommendation_Agent": Agent(
            role='Model_Recommendation_Agent',
            goal="Suggest the most suitable machine learning models.",
            backstory="Expert in recommending machine learning algorithms.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        "Starter_Code_Generator_Agent": Agent(
            role='Starter_Code_Generator_Agent',
            goal=f"Generate starter Python code for the project. Always give dataset name as {file_name}",
            backstory="Code wizard for generating starter code templates.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        "Code_Modification_Agent": Agent(
            role='Code_Modification_Agent',
            goal="Modify the generated Python code based on user suggestions.",
            backstory="Expert in adapting code according to user feedback.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        # "Code_Runner_Agent": Agent(
        #     role='Code_Runner_Agent',
        #     goal="Run the generated Python code and catch any errors.",
        #     backstory="Debugging expert.",
        #     verbose=True,
        #     allow_delegation=True,
        #     llm=llm,
        # ),
        "Code_Debugger_Agent": Agent(
            role='Code_Debugger_Agent',
            goal="Debug the generated Python code.",
            backstory="Seasoned code debugger.",
            verbose=True,
            allow_delegation=False,
            llm=llm,
        ),
        "Compiler_Agent":Agent(
            role = "Code_compiler",
            goal = "Extract only the python code.",
            backstory = "You are the compiler which extract only the python code.",
            verbose = True,
            allow_delegation = False,
            llm = llm
        )
    }

# Function to create tasks based on user inputs
def create_tasks(func_call,user_question,file_name, data_upload, df, suggestion, edited_code, debugger, agents):
    info = df.info()
    tasks = []
    if(func_call == "Process"):
        tasks.append(Task(
                description=f"Clarify the ML problem: {user_question}",
                agent=agents["Problem_Definition_Agent"],
                expected_output="A clear and concise definition of the ML problem."
            )
            )
        
        if data_upload:
            tasks.extend([
                Task(
                    description=f"Evaluate the data provided by the file name . This is the data: {df}",
                    agent=agents["EDA_Agent"],
                    expected_output="An assessment of the EDA and preprocessing like dataset info, missing value, duplication, outliers etc. on the data provided"
                ),
                Task(
                    description=f"Feature Engineering on data {df} based on EDA output: {info}",
                    agent=agents["Feature_Engineering_Agent"],
                    expected_output="An assessment of the Featuring Engineering and preprocessing like handling missing values, handling duplication, handling outliers, feature encoding, feature scaling etc. on the data provided"
                )
            ])

        tasks.extend([
            Task(
                description="Suggest suitable ML models.",
                agent=agents["Model_Recommendation_Agent"],
                expected_output="A list of suitable ML models."
            ),
            Task(
                description=f"Generate starter Python code based on feature engineering, where column names are {df.columns.tolist()}. Generate only the code without any extra text",
                agent=agents["Starter_Code_Generator_Agent"],
                expected_output="Starter Python code."
            ),
        ])
    if(func_call == "Modify"):
        if suggestion:
            tasks.append(
                Task(
                    description=f"Modify the already generated code {edited_code} according to the suggestion: {suggestion} \n\n Do not generate entire new code.",
                    agent=agents["Code_Modification_Agent"],
                    expected_output="Modified code."
                )
            )
    if(func_call == "Debug"):
        if debugger:
            tasks.append(
                Task(
                    description=f"Debug and fix any errors for data with column names {df.columns.tolist()} with data as {df} in the generated code: {edited_code}  \n\n According to the debugging: {debugger}. \n\n Do not generate entire new code. Just remove the error in the code by modifying only necessary parts of the code.",
                    agent=agents["Code_Debugger_Agent"],
                    expected_output="Debugged and successfully executed code."
                )
            )
    tasks.append(
        Task(
            description = "Your job is to only extract python code from string",
            agent = agents["Compiler_Agent"],
            expected_output = "Running python code."
        )
    )

    return tasks


main()