app.py

import gradio as gr
from crewai import Agent as CrewAgent, Task, Crew
import autogen
from huggingface_hub import InferenceClient
import os
import re
import numpy as np
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB
import asyncio

# Initialize the client with the Mistral-7B-Instruct-v0.2 model
client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.2")

# Shared context for both agents
SHARED_CONTEXT = """You are part of a multi-agent system designed to provide respectful, empathetic, and accurate support for Zerodha, a leading Indian financial services company. Your role is crucial in ensuring all interactions uphold the highest standards of customer service while maintaining Zerodha's excellent reputation.

Key points about Zerodha:
1. India's largest discount broker, known for innovative technology and low-cost trading.
2. Flat fee structure: ₹20 per executed order for intraday and F&O trades, zero brokerage for delivery equity investments.
3. Main trading platform: Kite (web and mobile).
4. Coin platform for commission-free direct mutual fund investments.
5. Extensive educational resources through Varsity.
6. Additional tools: Sentinel (price alerts) and ChartIQ (advanced charting).
7. Console for account management and administrative tasks.

Always prioritize user safety, ethical investing practices, and transparent communication. Never provide information that could mislead users or bring disrepute to Zerodha."""

# Guardrail functions (unchanged)
def sanitize_input(input_text):
    sanitized = re.sub(r'[<>&\']', '', input_text)
    return sanitized

approved_topics = ['account opening', 'trading', 'fees', 'platforms', 'funds', 'regulations', 'support']
vectorizer = CountVectorizer()
classifier = MultinomialNB()

X = vectorizer.fit_transform(approved_topics)
y = np.arange(len(approved_topics))
classifier.fit(X, y)

def is_relevant_topic(query):
    query_vector = vectorizer.transform([query])
    prediction = classifier.predict(query_vector)
    return prediction[0] in range(len(approved_topics))

def redact_sensitive_info(text):
    text = re.sub(r'\b\d{10,12}\b', '[REDACTED]', text)
    text = re.sub(r'[A-Z]{5}[0-9]{4}[A-Z]', '[REDACTED]', text)
    return text

def check_response_content(response):
    unauthorized_patterns = [
        r'\b(guarantee|assured|certain)\b.*\b(returns|profit)\b',
        r'\b(buy|sell)\b.*\b(specific stocks?|shares?)\b'
    ]
    return not any(re.search(pattern, response, re.IGNORECASE) for pattern in unauthorized_patterns)

def check_confidence(response):
    uncertain_phrases = ['I'm not sure', 'It's possible', 'I don't have enough information']
    return not any(phrase in response for phrase in uncertain_phrases)

def generate_response(prompt):
    response = client.text_generation(prompt, max_new_tokens=500, temperature=0.7)
    return response

def post_process_response(response):
    response = re.sub(r'\b(stupid|dumb|idiotic|foolish)\b', 'mistaken', response, flags=re.IGNORECASE)
    
    if not re.search(r'(Thank you|Is there anything else|Hope this helps|Let me know if you need more information)\s*$', response, re.IGNORECASE):
        response += "\n\nIs there anything else I can help you with regarding Zerodha's services?"
    
    if re.search(r'\b(invest|trade|buy|sell|market)\b', response, re.IGNORECASE):
        response += "\n\nPlease note that this information is for educational purposes only and should not be considered as financial advice. Always do your own research and consider consulting with a qualified financial advisor before making investment decisions."
    
    return response

# CrewAI and AutoGen setup (unchanged)
communication_expert_crew = CrewAgent(
    role='Communication Expert',
    goal='Interpret and rephrase user queries with empathy and respect',
    backstory="""You are an expert in communication, specializing in understanding and rephrasing queries to ensure they are interpreted in the most positive and constructive light. Your role is crucial in setting the tone for respectful and empathetic interactions.""",
    verbose=True,
    allow_delegation=False,
    tools=[generate_response]
)

response_expert_crew = CrewAgent(
    role='Response Expert',
    goal='Provide accurate, helpful, and emotionally intelligent responses to user queries',
    backstory="""You are an expert in Zerodha's services and policies, with a keen ability to provide comprehensive and empathetic responses. Your role is to ensure that all user queries are addressed accurately while maintaining a respectful and supportive tone.""",
    verbose=True,
    allow_delegation=False,
    tools=[generate_response]
)

communication_expert_autogen = autogen.AssistantAgent(
    name="Communication_Expert",
    system_message=SHARED_CONTEXT + """
As the Communication Expert, your primary role is to interpret user queries with the utmost respect and empathy. You should:
1. Rephrase the user's query to ensure it's understood in the most positive and constructive light.
2. Identify and highlight any emotional subtext or concerns in the query.
3. Frame the query in a way that invites a supportive and informative response.
4. Ensure that any potential complaints or frustrations are acknowledged respectfully.

Your output should be a rephrased version of the user's query that maintains its original intent while setting the stage for an empathetic and respectful response.""",
    llm_config={"config_list": [{"model": "gpt-3.5-turbo"}]}
)

response_expert_autogen = autogen.AssistantAgent(
    name="Response_Expert",
    system_message=SHARED_CONTEXT + """
As the Response Expert, your role is to provide accurate, helpful, and emotionally intelligent responses to user queries. You should:
1. Address the user's question or concern directly and comprehensively.
2. Maintain a tone of respect and empathy throughout your response.
3. Provide clear, factual information about Zerodha's services and policies.
4. When discussing financial matters, include appropriate disclaimers and encourage users to seek professional advice for complex decisions.
5. For complaints or concerns, acknowledge them respectfully and provide constructive guidance or escalation paths.
6. Always uphold Zerodha's reputation for transparency and user-centric service.

Your output should be a complete, informative response that addresses the user's query while demonstrating empathy and respect.""",
    llm_config={"config_list": [{"model": "gpt-3.5-turbo"}]}
)

user_proxy = autogen.UserProxyAgent(
    name="User_Proxy",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=1
)

# Main function rewritten to be asynchronous
async def zerodha_support(message, history):
    sanitized_message = sanitize_input(message)
    
    if not is_relevant_topic(sanitized_message):
        return "I'm sorry, but I can only assist with queries related to Zerodha's services and trading. Could you please ask a question about your Zerodha account, trading, or our platforms?"
    
    sanitized_message = redact_sensitive_info(sanitized_message)

    # Use crewAI for initial query rephrasing
    rephrase_task = Task(
        description=f"Rephrase the following user query with empathy and respect: '{sanitized_message}'",
        agent=communication_expert_crew
    )

    crew = Crew(
        agents=[communication_expert_crew],
        tasks=[rephrase_task],
        verbose=2
    )

    rephrased_query = crew.kickoff()

    # Use AutoGen for generating the response
    async def get_autogen_response():
        await user_proxy.a_initiate_chat(
            response_expert_autogen,
            message=f"Please provide a respectful and empathetic response to the following query: '{rephrased_query}'"
        )
        return response_expert_autogen.last_message()["content"]

    response = await get_autogen_response()

    if not check_response_content(response):
        response += "\n\nPlease note that I cannot provide specific investment advice or guarantee returns. For personalized guidance, please consult with a qualified financial advisor."

    if not check_confidence(response):
        return "I apologize, but I'm not confident in providing an accurate answer to this query. For the most up-to-date and accurate information, please contact Zerodha's customer support directly."

    final_response = post_process_response(response)

    return final_response

# Wrap the asynchronous function for Gradio
def zerodha_support_wrapper(message, history):
    return asyncio.run(zerodha_support(message, history))

# Gradio interface setup
demo = gr.ChatInterface(
    zerodha_support_wrapper,
    chatbot=gr.Chatbot(height=600),
    textbox=gr.Textbox(placeholder="Ask your question about Zerodha here...", container=False, scale=7),
    title="Zerodha Support Assistant",
    description="Ask questions about Zerodha's services, trading, account management, and more. Our multi-agent system ensures respectful and empathetic responses.",
    theme="soft",
    examples=[
        "How do I open a Zerodha account?",
        "I'm frustrated with the recent changes to the Kite platform. Can you help?",
        "What are the risks involved in F&O trading?",
        "I think there's an error in my account statement. What should I do?",
        "Can you explain Zerodha's policy on intraday trading margins?",
        "I'm new to investing. What resources does Zerodha offer for beginners?",
        "How does Zerodha ensure the security of my investments and personal data?"
    ],
)

if __name__ == "__main__":
    public_url = demo.launch(share=True, server_name="0.0.0.0", server_port=7860)
    print(f"\n\nSHAREABLE LINK: {public_url}\n\n")