app.py

import gradio as gr
from transformers import pipeline
import pandas as pd
import spaces
import plotly.express as px

# Load dataset
from datasets import load_dataset
ds = load_dataset('ZennyKenny/demo_customer_nps')
df = pd.DataFrame(ds['train'])

# Initialize model pipeline
from huggingface_hub import login
import os

# Login using the API key stored as an environment variable
hf_api_key = os.getenv("API_KEY")
login(token=hf_api_key)

classifier = pipeline("text-classification", model="distilbert/distilbert-base-uncased-finetuned-sst-2-english")
generator = pipeline("text2text-generation", model="google/flan-t5-base")

# Function to classify customer comments
@spaces.GPU
def classify_comments(categories):
    global df  # Ensure we're modifying the global DataFrame
    sentiments = []
    assigned_categories = []
    for comment in df['customer_comment']:
        # Classify sentiment
        sentiment = classifier(comment)[0]['label']
        # Generate category
        category_str = ', '.join(categories)
        prompt = f"What category best describes this comment? '{comment}' Please answer using only the name of the category: {category_str}."
        category = generator(prompt, max_length=30)[0]['generated_text']
        assigned_categories.append(category)
        sentiments.append(sentiment)
    df['comment_sentiment'] = sentiments
    df['comment_category'] = assigned_categories
    return df[['customer_id', 'customer_comment', 'comment_sentiment', 'comment_category', 'customer_nps', 'customer_segment']].to_html(index=False)

# Function to generate visualizations
def visualize_output():
    # Ensure the required columns exist
    if 'comment_sentiment' not in df.columns or 'comment_category' not in df.columns:
        # Return 5 values (None for plots and an error message for markdown)
        return None, None, None, "Error: Please classify comments before visualizing.", None

    # Pie Chart of Sentiment
    sentiment_counts = df['comment_sentiment'].value_counts()
    sentiment_pie = px.pie(
        values=sentiment_counts.values,
        names=sentiment_counts.index,
        title="Sentiment Distribution",
        hover_data=[sentiment_counts.values],
        labels={'value': 'Count', 'names': 'Sentiment'}
    )
    sentiment_pie.update_traces(textinfo='percent+label', hovertemplate="Sentiment: %{label}<br>Count: %{value}<br>Percentage: %{percent}")

    # Pie Chart of Comment Categories
    category_counts = df['comment_category'].value_counts()
    category_pie = px.pie(
        values=category_counts.values,
        names=category_counts.index,
        title="Comment Category Distribution",
        hover_data=[category_counts.values],
        labels={'value': 'Count', 'names': 'Category'}
    )
    category_pie.update_traces(textinfo='percent+label', hovertemplate="Category: %{label}<br>Count: %{value}<br>Percentage: %{percent}")

    # Stacked Bar Chart of Sentiment by Category
    sentiment_by_category = df.groupby(['comment_category', 'comment_sentiment']).size().unstack()
    stacked_bar = px.bar(
        sentiment_by_category,
        barmode='stack',
        title="Sentiment by Comment Category",
        labels={'value': 'Count', 'comment_category': 'Category', 'comment_sentiment': 'Sentiment'}
    )

    # KPI Visualizations
    avg_nps = df['customer_nps'].mean()
    avg_nps_positive = df[df['comment_sentiment'] == 'POSITIVE']['customer_nps'].mean()
    avg_nps_negative = df[df['comment_sentiment'] == 'NEGATIVE']['customer_nps'].mean()
    avg_nps_by_category = df.groupby('comment_category')['customer_nps'].mean().reset_index()
    avg_nps_by_segment = df.groupby('customer_segment')['customer_nps'].mean().reset_index()

    kpi_visualization = f"""
    **Average NPS Scores:**
    - Overall: {avg_nps:.2f}
    - Positive Sentiment: {avg_nps_positive:.2f}
    - Negative Sentiment: {avg_nps_negative:.2f}
    **Average NPS by Category:**
    {avg_nps_by_category.to_markdown(index=False)}
    **Average NPS by Segment:**
    {avg_nps_by_segment.to_markdown(index=False)}
    """

    # Pie Chart of Sentiment by Customer Segment
    sentiment_by_segment = df.groupby(['customer_segment', 'comment_sentiment']).size().unstack()
    sentiment_by_segment_pie = px.pie(
        sentiment_by_segment,
        title="Sentiment by Customer Segment",
        labels={'value': 'Count', 'customer_segment': 'Segment', 'comment_sentiment': 'Sentiment'}
    )

    return sentiment_pie, category_pie, stacked_bar, kpi_visualization, sentiment_by_segment_pie

# Gradio Interface
with gr.Blocks() as nps:
    # State to store categories
    categories = gr.State([])

    # Function to add a category
    def add_category(categories, new_category):
        if new_category.strip() != "" and len(categories) < 5:  # Limit to 5 categories
            categories.append(new_category.strip())
        return categories, "", f"**Categories:**\n" + "\n".join([f"- {cat}" for cat in categories])

    # Function to reset categories
    def reset_categories():
        return [], "**Categories:**\n- None"

    # UI for adding categories
    with gr.Row():
        category_input = gr.Textbox(label="New Category", placeholder="Enter category name")
        add_category_btn = gr.Button("Add Category")
        reset_btn = gr.Button("Reset Categories")
        category_status = gr.Markdown("**Categories:**\n- None")

    # File upload and template buttons
    uploaded_file = gr.File(label="Upload CSV", type="filepath")
    template_btn = gr.Button("Use Template")
    gr.Markdown("# NPS Comment Categorization")

    # Classify button
    classify_btn = gr.Button("Classify Comments")
    output = gr.HTML()

    # Visualize button
    visualize_btn = gr.Button("Visualize Output")
    sentiment_pie = gr.Plot(label="Sentiment Distribution")
    category_pie = gr.Plot(label="Comment Category Distribution")
    stacked_bar = gr.Plot(label="Sentiment by Comment Category")
    kpi_visualization = gr.Markdown()
    sentiment_by_segment_pie = gr.Plot(label="Sentiment by Customer Segment")

    # Function to load data from uploaded CSV
    def load_data(file):
        global df  # Ensure we're modifying the global DataFrame
        if file is not None:
            file.seek(0)  # Reset file pointer
            if file.name.endswith('.csv'):
                custom_df = pd.read_csv(file, encoding='utf-8')
            else:
                return "Error: Uploaded file is not a CSV."
            # Check for required columns
            required_columns = ['customer_id', 'customer_comment', 'customer_nps', 'customer_segment']
            if not all(col in custom_df.columns for col in required_columns):
                return f"Error: Uploaded CSV must contain the following columns: {', '.join(required_columns)}"
            df = custom_df
            return "Custom CSV loaded successfully!"
        else:
            return "No file uploaded."

    # Function to use template categories
    def use_template():
        template_categories = ["Product Experience", "Customer Support", "Price of Service", "Other"]
        return template_categories, f"**Categories:**\n" + "\n".join([f"- {cat}" for cat in template_categories])

    # Event handlers
    add_category_btn.click(
        fn=add_category,
        inputs=[categories, category_input],
        outputs=[categories, category_input, category_status]
    )
    reset_btn.click(
        fn=reset_categories,
        outputs=[categories, category_status]
    )
    uploaded_file.change(
        fn=load_data,
        inputs=uploaded_file,
        outputs=output
    )
    template_btn.click(
        fn=use_template,
        outputs=[categories, category_status]
    )
    classify_btn.click(
        fn=classify_comments,
        inputs=categories,
        outputs=output
    )
    visualize_btn.click(
        fn=visualize_output,
        outputs=[sentiment_pie, category_pie, stacked_bar, kpi_visualization, sentiment_by_segment_pie]
    )

nps.launch(share=True)