app.py

from sentence_transformers import SentenceTransformer, util, CrossEncoder
from datasets import load_dataset
import pandas as pd
import torch

#Get the netflix dataset
netflix = load_dataset('hugginglearners/netflix-shows',use_auth_token=True)

#Filter for relevant columns and convert to pandas
netflix_df = netflix['train'].to_pandas()
netflix_df = netflix_df[['type','title','country','cast','release_year','rating','duration','listed_in','description']]

#load mpnet model
model = SentenceTransformer('all-mpnet-base-v2')

#load embeddings
flix_ds = load_dataset("nickmuchi/netflix-shows-mpnet-embeddings", use_auth_token=True)
dataset_embeddings = torch.from_numpy(flix_ds["train"].to_pandas().to_numpy()).to(torch.float)

#load cross-encoder for reranking
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-12-v2')

def display_df_as_table(model,top_k,score='score'):
    # Display the df with text and scores as a table
    df = pd.DataFrame([(hit[score],passages[hit['corpus_id']]) for hit in model[0:top_k]],columns=['Score','Text'])
    df['Score'] = round(df['Score'],2)
    df = df.merge(netflix_df,how='inner',left_on='Text',right_on='description')
    df.drop('Text',inplace=True,axis=1)
    
    return df

#function for generating similarity of query and netflix shows
def semantic_search(query,embeddings,top_k=top_k):
    '''Encode query and check similarity with embeddings'''
    
    question_embedding = model.encode(query, convert_to_tensor=True).cpu()
    hits = util.semantic_search(question_embedding, embeddings, top_k=top_k)
    hits = hits[0]
    
    ##### Re-Ranking #####
    # Now, score all retrieved passages with the cross_encoder
    cross_inp = [[query, netflix_df['description'].iloc[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)
    
    # Sort results by the cross-encoder scores
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]
    
    #Bi-encoder df   
    hits = sorted(hits, key=lambda x: x['score'], reverse=True)
    bi_df = display_df_as_table(hits,top_k)
    
    #Cross encoder df
    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
    cross_df = display_df_as_table(hits,top_k,'cross-score')
    
    return bi_df, cross_df
    
    
title = """<h1 id="title">Netflix Shows Semantic Search</h1>"""

description = """
Semantic Search is a way to generate search results based on the actual meaning of the query instead of a standard keyword search. I believe this way of searching provides more meaning results when trying to find a good show to watch on Netflix. For example, one could search for "Success, rags to riches story" as provided in the example below to generate shows or movies with a description that is semantically similar to the query. 

- The App generates embeddings using [All-Mpnet-Base-v2](https://huggingface.co/sentence-transformers/all-mpnet-base-v2) model from Sentence Transformers. 
- The model encodes the query and the discerption field from the [Netflix-Shows](https://huggingface.co/datasets/hugginglearners/netflix-shows) dataset which contains 8800 shows and movies currently on Netflix scraped from the web using Selenium. 
- Similarity scores are then generated, from highest to lowest. The user can select how many suggestions they need from the results.
- A Cross Encoder then re-ranks the top selections to further improve on the similarity scores.
- You will see 2 tables generated, one from the bi-encoder and the other from the cross encoder which further enhances the similarity score rankings

Enjoy and Search like you mean it!! 
"""
example_queries = ["Success, rags to riches","murder, crime scene investigation thriller"]

twitter_link = """
[![](https://img.shields.io/twitter/follow/nickmuchi?label=@nickmuchi&style=social)](https://twitter.com/nickmuchi)
"""

css = '''
h1#title {
  text-align: center;
}
'''

demo = gr.Blocks(css=css)

with demo:
    gr.Markdown(title)
    gr.Markdown(description)
    gr.Markdown(twitter_link)
    
    top_k = gr.Slider(minimum=3,maximum=10,value=5,step=1,label='Number of Suggestions to Generate')
    

    with gr.Row():
        query = gr.Textbox(lines=3,label='Describe the Netflix show or movie you would like to watch..')
            
    with gr.Row():
        gr.Markdown(f'''Top-{top_k} Bi-Encoder Retrieval hits''')
        bi_output = gr.DataFrame(headers=['Similarity Score','Type','Title','Country','Cast','Release Year','Rating','Duration','Category Listing','Description'])
        
    with gr.Row():
        gr.Markdown(f'''Top-{top_k} Cross-Encoder Re-ranker hits''')
        cross_output = gr.DataFrame(headers=['Similarity Score','Type','Title','Country','Cast','Release Year','Rating','Duration','Category Listing','Description'])
        
    with gr.Row():
        example_url = gr.Examples(examples=example_queries,inputs=[query])
        
    
    sem_but = gr.Button('Search')
     
            
    sem_but.click(semantic_search,inputs=[query,dataset_embeddings,img_input,top_k],outputs=[bi_output,cross_output],queue=True)

    gr.Markdown("![visitor badge](https://visitor-badge.glitch.me/badge?page_id=nickmuchi-netflix-shows-semantic-search)")

    
demo.launch(debug=True,enable_queue=True)