app.py

import pandas as pd
from sentence_transformers import SentenceTransformer, util
from transformers import pipeline
import torch
import gradio as gr
import os 

# Use the relative path where the CSV is uploaded
csv_file_path = os.path.join(os.getcwd(), 'analytics_vidhya_data.csv')

# Load the dataset
df = pd.read_csv(csv_file_path, encoding='ISO-8859-1')

# Load the pre-trained model for embeddings (using SentenceTransformers)
model = SentenceTransformer('multi-qa-mpnet-base-dot-v1')

# Combine title and description to create a full text for each course
df['full_text'] = df.iloc[:,0] + " " + df.iloc[:,1] + " " + df['Instructor Name'] + " " + str(df['Rating']) + " " + df['Category']

# Convert full course texts into embeddings
course_embeddings = model.encode(df['full_text'].tolist(), convert_to_tensor=True)

# Function to expand the query using paraphrasing
def expand_query(query):
    paraphraser = pipeline('text2text-generation', model='Vamsi/T5_Paraphrase_Paws')
    expanded_queries = paraphraser(query, num_return_sequences=3, max_length=50, do_sample=True)
    return [q['generated_text'] for q in expanded_queries]

# Function to search for the most relevant courses
def search_courses(query, level_filter=None, category_filter=None, top_k=3):
    # Step 1: Expand the query using paraphrasing
    expanded_queries = expand_query(query)

    # Step 2: Initialize an array to store all similarities
    all_similarities = []

    for expanded_query in expanded_queries:
        # Convert each expanded query into an embedding
        query_embedding = model.encode(expanded_query, convert_to_tensor=True)

        # Compute cosine similarities between the query embedding and course embeddings
        similarities = util.pytorch_cos_sim(query_embedding, course_embeddings)[0]

        # Append to the list of all similarities
        all_similarities.append(similarities)

    # Step 3: Convert the list of tensors to a single tensor by taking the maximum similarity for each course
    aggregated_similarities = torch.max(torch.stack(all_similarities), dim=0)[0]

    # Step 4: Apply filters
    filtered_df = df.copy()
    if level_filter:
        filtered_df = filtered_df[filtered_df['Level of Difficulty'] == level_filter]
    if category_filter:
        filtered_df = filtered_df[filtered_df['Category'] == category_filter]
    
    if filtered_df.empty:
        return "<p>No matching courses found.</p>"
    
    # Recalculate similarities for the filtered data
    filtered_similarities = aggregated_similarities[filtered_df.index]

    # Step 5: Get top_k most similar courses
    top_results = filtered_similarities.topk(k=min(top_k, len(filtered_similarities)))

    # Prepare the output as clickable links
    results = []
    for idx in top_results.indices:
        idx = int(idx)
        course_title = filtered_df.iloc[idx]['Course Title']
        course_description = filtered_df.iloc[idx,1]
        course_url = filtered_df.iloc[idx,-1]
        
        
        # Format the result as a clickable hyperlink using raw HTML
        course_link = f'<a href="{course_url}" target="_blank">{course_title}</a>'
        results.append(f"<strong>{course_link}</strong><br>{course_description}<br><br>")

    # Combine all results into an HTML formatted list
    return "<ol>" + "".join([f"<li>{result}</li>" for result in results]) + "</ol>"

# Create Gradio UI
def create_gradio_interface():
    with gr.Blocks() as demo:
        gr.Markdown("# 📚 Analytics Vidhya Free Courses")
        gr.Markdown("Enter your query and use filters to narrow down the search.")

        # Input elements
        query = gr.Textbox(label="🔍 Search for a course", placeholder="Enter course topic or description")

        # Filters (in a collapsible form)
        with gr.Accordion("🔍 Filters", open=False):
            level_filter = gr.Dropdown(choices=["Beginner", "Intermediate", "Advanced"], label="📚 Course Level", multiselect=False)
            category_filter = gr.Dropdown(choices=["Data Science", "Machine Learning", "Deep Learning", "AI", "NLP"], label="📂 Category", multiselect=False)

        # Search button
        search_button = gr.Button("Search")

        # Output HTML for displaying results
        output = gr.HTML(label="Search Results")

        # On button click, trigger the search function
        search_button.click(fn=search_courses, inputs=[query, level_filter, category_filter], outputs=output)

    return demo

# Launch Gradio interface
demo = create_gradio_interface()
demo.launch(share=True, debug=True)