import gradio as gr import random from huggingface_hub import InferenceClient from sentence_transformers import SentenceTransformer import torch theme = gr.themes.Base() # use a simple base theme, we'll add custom CSS #spotify code spotify_embed_code = """ """ theme = gr.themes.Ocean( primary_hue="green", secondary_hue="cyan", neutral_hue="sky", ).set( body_background_fill ="f1eeccff", button_secondary_background_fill="#ABA8A6", button_primary_text_color="#004643", background_fill_primary="neutral_200" ) gradient_css = """ body { background: linear-gradient(135deg, #76c893, #1a759f); } """ # SEMANTIC SEARCH STEP 2 with open("songs_knowledge_base_2.txt", "r", encoding="utf-8") as file: songs_knowledge_base_text = file.read() # SEMANTIC SEARCH STEP 3 def preprocess_text(text): # Strip extra whitespace from the beginning and the end of the text cleaned_text = text.strip() # Split the cleaned_text by every newline character (\n) chunks = cleaned_text.split("*") # Create an empty list to store cleaned chunks cleaned_chunks = [] # Write your for-in loop below to clean each chunk and add it to the cleaned_chunks list for chunk in chunks: stripped_chunk = chunk.strip() cleaned_chunks.append(stripped_chunk) # Print cleaned_chunks print(cleaned_chunks) # Print the length of cleaned_chunks print(len(cleaned_chunks)) # Return the cleaned_chunks return cleaned_chunks # Call the preprocess_text function and store the result in a cleaned_chunks variable cleaned_chunks = preprocess_text(songs_knowledge_base_text) #SEMANTIC STEP 4 # Load the pre-trained embedding model that converts text to vectors model = SentenceTransformer('all-MiniLM-L6-v2') def create_embeddings(text_chunks): # Convert each text chunk into a vector embedding and store as a tensor chunk_embeddings = model.encode(text_chunks, convert_to_tensor=True) # Replace ... with the text_chunks list #contains number version of words that we put in chunks # Print the chunk embeddings print(chunk_embeddings) # Print the shape of chunk_embeddings print(chunk_embeddings.shape) # Return the chunk_embeddings return chunk_embeddings # Call the create_embeddings function and store the result in a new chunk_embeddings variable chunk_embeddings = create_embeddings(cleaned_chunks) #SEMANTICS step 5 # Define a function to find the most relevant text chunks for a given query, chunk_embeddings, and text_chunks def get_top_chunks(query, chunk_embeddings, text_chunks): # Convert the query text into a vector embedding query_embedding = model.encode(query, convert_to_tensor=True) # Complete this line # Normalize the query embedding to unit length for accurate similarity comparison query_embedding_normalized = query_embedding / query_embedding.norm() # Normalize all chunk embeddings to unit length for consistent comparison chunk_embeddings_normalized = chunk_embeddings / chunk_embeddings.norm(dim=1, keepdim=True) # Calculate cosine similarity between query and all chunks using matrix multiplication similarities = torch.matmul(chunk_embeddings_normalized, query_embedding_normalized) # Complete this line # Print the similarities print(similarities) # Find the indices of the 3 chunks with highest similarity scores top_indices = torch.topk(similarities, k=10).indices #gives index values of top chunks # Print the top indices print(top_indices) # Create an empty list to store the most relevant chunks top_chunks = [] # Loop through the top indices and retrieve the corresponding text chunks for i in top_indices: relevant_info = cleaned_chunks[i] top_chunks.append(relevant_info) # stored the text version of the index chunks we got from top_indcies # Return the list of most relevant chunks return top_chunks client = InferenceClient('Qwen/Qwen2.5-72B-Instruct') def respond(message, history): info = get_top_chunks(message, chunk_embeddings, cleaned_chunks) messages = [{'role': 'system', 'content': f'You are a friendly chatbot using {info} to answer questions. You love creating playlists and will give at least 10 songs as a response. You will also capitalize the first letters of the first and last names of every artist you name. Also very important, make sure the songs match pretty accurately what the user is asking based on factors like tempo, emotion, lyrics, and more'}] if history: messages.extend(history) messages.append({"role": "user", "content": message}) response = client.chat_completion(messages, max_tokens = 600) return response["choices"][0]["message"]["content"].strip() def echo(message, history): return message def yes_no(message, history): responses = ["Yes", "No"] return random.choice(responses) with gr.Blocks(theme=theme, css=gradient_css) as chatbot: with gr.Row(scale=1): with gr.Column(scale=4): gr.Image("music_banner_pic.png") #width=10000, height=300) with gr.Row(scale=3): with gr.Column(scale=1): gr.Image("White_AuxAI_logo .png") with gr.Column(scale=4): gr.ChatInterface(respond, type="messages", title = "AuxAI", theme = theme, description = "Hi! I’m AuxAI, your friendly music recommendation assistant. Tell me your mood, genre, or style and I’ll suggest some songs! AuxAI is on a mission to make music discovery personalized, inspiring and effortless. Our software connects the user to a playlist of their liking that perfectly match their mood, vibe or moment. We hope you uncover favorites and rediscover old ones, turning each search into a seamless find!", examples=["Give me some jazzy songs to study and focus to", "Give me a playlist of hype, upbeat songs to workout to", "Give me songs like Espresso by Sabrina Carpenter"],) #__________________________________________ #ADD SONG PLAYLIST HERE, EXAMPLE IMAGES, AND LINKS TO RESOURCES with gr.Row(scale=1): gr.Markdown("### Enjoy some of our favorite songs!") with gr.Row(scale=1): gr.HTML(spotify_embed_code) # with gr.Row(scale=1): # with gr.Column(): # resources links here #__________________________________________ chatbot.launch()