app.py

import os
import torch
import gradio as gr
from langchain import HuggingFaceHub
from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate
from transformers import pipeline




# 2. Model Initialization
# Initialize sentiment analyzer
sentiment_analyzer = pipeline(
    "sentiment-analysis",
    model="finiteautomata/bertweet-base-sentiment-analysis",
    device=0 if torch.cuda.is_available() else -1
)

# Initialize LLM
llm = HuggingFaceHub(
    repo_id="deepseek-ai/deepseek-coder-33b-instruct",
    model_kwargs={"temperature": 0.7}
)

# 3. Templates
email_template = PromptTemplate(
    input_variables=["previous_interaction", "situation_type", "tone", "urgency"],
    template="""Based on these details, generate a professional follow-up email:

Previous Interaction: {previous_interaction}
Situation Type: {situation_type}
Tone: {tone}
Urgency Level: {urgency}

Generate a personalized email that:
1. Maintains {tone} tone
2. Addresses the specific situation
3. Provides clear next steps
4. Is appropriate for {urgency} urgency level
"""
)

scoring_template = """
Analyze this follow-up email carefully and provide scores on a scale of 1-10 for each category:

Email to analyze:
{email_text}

Please provide numerical scores and explanations in this exact format:

CLARITY SCORE: [1-10]
Explanation: [Why this score]

PROFESSIONALISM SCORE: [1-10]
Explanation: [Why this score]

ACTION ITEMS SCORE: [1-10]
Explanation: [Why this score]

PERSONALIZATION SCORE: [1-10]
Explanation: [Why this score]

OVERALL EFFECTIVENESS SCORE: [1-10]
Explanation: [Why this score]

IMPROVEMENT SUGGESTIONS:
1. [First suggestion]
2. [Second suggestion]
3. [Third suggestion]
"""

# 4. Create LangChain
email_chain = LLMChain(llm=llm, prompt=email_template)

# 5. Helper Functions
def analyze_sentiment(text):
    try:
        result = sentiment_analyzer(text)[0]
        sentiment_to_tone = {
            'POS': 'Friendly',
            'NEU': 'Professional',
            'NEG': 'Apologetic'
        }
        return sentiment_to_tone.get(result['label'], 'Professional')
    except Exception as e:
        return 'Professional'

# 6. Main Generation Function
def generate_followup_email(previous_interaction, situation_type, tone, urgency):
    try:
        if not tone:
            tone = analyze_sentiment(previous_interaction)
            
        # Generate email
        email_result = email_chain.run({
            "previous_interaction": previous_interaction,
            "situation_type": situation_type,
            "tone": tone,
            "urgency": urgency
        })
        
        # Generate score
        score_result = llm(scoring_template.format(email_text=email_result))
        return email_result, score_result
    except Exception as e:
        return f"Error generating email: {str(e)}", "Scoring unavailable"

# 7. Gradio Interface
demo = gr.Interface(
    fn=generate_followup_email,
    inputs=[
        gr.Textbox(
            label="Previous Interaction",
            lines=5,
            placeholder="Describe the previous interaction with the customer..."
        ),
        gr.Dropdown(
            label="Situation Type",
            choices=[
                "Complaint Resolution",
                "Service Issue",
                "Payment Dispute",
                "Product Query",
                "General Follow-up"
            ]
        ),
        gr.Dropdown(
            label="Tone (Optional - will be automatically detected if not specified)",
            choices=[
                "",
                "Professional",
                "Apologetic",
                "Friendly",
                "Formal",
                "Empathetic"
            ]
        ),
        gr.Dropdown(
            label="Urgency",
            choices=["High", "Medium", "Low"]
        )
    ],
    outputs=[
        gr.Textbox(label="Generated Email"),
        gr.Textbox(label="Email Score and Suggestions")
    ],
    title="Smart Sales Email Generator with Quality Scoring",
    description="Generate and evaluate follow-up emails based on previous interactions",
    examples=[
        [
            "Customer complained about slow website loading times and threatened to cancel subscription",
            "Complaint Resolution",
            "Apologetic",
            "High"
        ],
        [
            "Client requested information about premium features and pricing",
            "Product Query",
            "Professional",
            "Medium"
        ]
    ]
)

# 8. Launch App
if __name__ == "__main__":
    demo.launch()