app.py

import gradio as gr
import asyncio
import random
import torch
from huggingface_hub import InferenceClient
from sentence_transformers import SentenceTransformer

# ---------------------- THEME ----------------------
theme = gr.themes.Ocean(
    secondary_hue="lime",
    neutral_hue="teal",
    text_size="lg",
    spacing_size="lg",
).set(
    body_background_fill='*primary_400',
    body_background_fill_dark='*primary_950',
    body_text_color='*primary_50',
    body_text_color_dark='*primary_50',
    background_fill_primary_dark='*secondary_500',
    background_fill_secondary='*primary_700',
    background_fill_secondary_dark='*primary_900',
    button_primary_background_fill='linear-gradient(120deg, *secondary_800 0%, *primary_300 60%, *primary_800 100%)',
    button_primary_background_fill_dark='linear-gradient(120deg, *secondary_400 0%, *primary_400 60%, *primary_600 100%)',
    button_primary_background_fill_hover='linear-gradient(120deg, *secondary_400 0%, *primary_300 60%, *neutral_300 100%)',

)

# ---------------------- LOAD KNOWLEDGE BASE ----------------------
with open("Skin_cancer_harvard.txt", "r", encoding="utf-8") as file:
    Skin_cancer_harvard_text = file.read()
print(Skin_cancer_harvard_text)

def preprocess_text(text):
    cleaned_text = text.strip()
    chunks = cleaned_text.split("\n")
    cleaned_chunks = [chunk.strip() for chunk in chunks if chunk.strip()]
    print(cleaned_chunks)
    print(len(cleaned_chunks))
    return cleaned_chunks

cleaned_chunks = preprocess_text(Skin_cancer_harvard_text)

# ---------------------- EMBEDDINGS ----------------------
model = SentenceTransformer('all-MiniLM-L6-v2')

def create_embeddings(text_chunks):
    chunk_embeddings = model.encode(text_chunks, convert_to_tensor=True)
    print(chunk_embeddings.shape)
    return chunk_embeddings

chunk_embeddings = create_embeddings(cleaned_chunks)

# ---------------------- SEMANTIC SEARCH ----------------------
def get_top_chunks(query, chunk_embeddings, text_chunks):
    query_embedding = model.encode(query, convert_to_tensor=True)
    query_embedding_normalized = query_embedding / query_embedding.norm()
    chunk_embeddings_normalized = chunk_embeddings / chunk_embeddings.norm(dim=1, keepdim=True)
    similarities = torch.matmul(chunk_embeddings_normalized, query_embedding_normalized)
    top_indices = torch.topk(similarities, k=3).indices
    top_chunks = [cleaned_chunks[i] for i in top_indices]
    return top_chunks

# ---------------------- LLM CLIENT ----------------------
client = InferenceClient("microsoft/phi-4")

# ---------------------- CHAT FUNCTION ----------------------
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 are always willing to help and want the best for the user. '
                'You need to emphasize that you are not a medical professional at the end '
                'of the message, but you are here to help to the best of your ability. '
                'Be confident and comforting to the users when helping them. '
                'In your response add suggestions for a couple follow-up questions '
                'to further the conversation with the chatbot.'
            )
        }
    ]

    if history:
        messages.extend(history)
    messages.append({'role': 'user', 'content': message})

    # Run blocking HF API in background thread (prevents StopIteration error)
    def blocking_call():
        return client.chat_completion(messages, max_tokens=500, top_p=0.8)

    response = asyncio.run(asyncio.to_thread(blocking_call))
    content = response['choices'][0]['message']['content'].strip()
    history.append((message, content))
    return history, content

# ---------------------- GRADIO APP ----------------------
with gr.Blocks(theme=theme) as chatbot:
    with gr.Row(scale=1):
        gr.Image("Capstone_Banner.png")

    with gr.Row():
        with gr.Column(scale=1):
            gr.Image("Aloe_the_Turtle.png")

        with gr.Row():
            gr.Markdown(
                "Click the button below to access the teachable machine, an AI Visual Scanner to detect Skin Cancer. "
                "The main purpose of this teachable machine is to check if you have a cancerous or non-cancerous mole. "
                "Place your mole near your camera and the analysis will be represented below. "
                "Note that these results are not 100% accurate, so be sure to consult a medical professional if you have any concerns."
            )

        with gr.Row(scale=1):
            gr.Button(
                value="AI Visual Testing Moles for Skin Cancer!",
                link="https://teachablemachine.withgoogle.com/models/onfoEa0p-/"
            )

        with gr.Column(scale=3):
            gr.ChatInterface(
                fn=respond,
                title="Your Personal Skin Chatbot!",
                description=(
                    "Welcome, my name is Aloe the Turtle and I am here to help you address any dermatology-related "
                    "questions you may have on topics such as Skin Cancer, Acne, Eczema, and much more. "
                    "Just remember, while I have comprehensive knowledge on skin concerns, I am not a medical professional!"
                ),
                type="messages",
                theme=theme,
                examples=[
                    "What ingredients should I use to clear my Acne?",
                    "What can I do to proactively prevent Skin Cancer?",
                    "How do I tell the difference between eczema and psoriasis?"
                ]
            )

chatbot.launch()