app.py

import gradio as gr
from transformers import pipeline
from PIL import Image
import torch
import os
import spaces

# --- Configuration & Model Loading ---
# Use the pipeline, which is more robust as seen in the working example
print("Loading MedGemma model via pipeline...")
model_loaded = False
pipe = None
try:
    pipe = pipeline(
        "image-to-text",
        model="google/medgemma-4b-it",
        model_kwargs={"torch_dtype": torch.bfloat16},
        device_map="auto",
        token=os.environ.get("HF_TOKEN")
    )
    model_loaded = True
    print("Model loaded successfully!")
except Exception as e:
    print(f"Error loading model: {e}")


# --- Core Chatbot Function ---
@spaces.GPU
def symptom_checker_chat(user_input, history_for_display, new_image_upload, image_state):
    """
    Manages the conversation using the correct message format derived from the working example.
    """
    if not model_loaded:
        if user_input:
            history_for_display.append((user_input, "Error: The model could not be loaded."))
        return history_for_display, image_state, None, ""

    current_image = new_image_upload if new_image_upload is not None else image_state

    # Build the 'messages' list using the correct format for the pipeline
    messages = []
    
    # Process the conversation history
    for user_msg, assistant_msg in history_for_display:
        messages.append({"role": "user", "content": user_msg})
        if assistant_msg:
            messages.append({"role": "assistant", "content": assistant_msg})

    # Add the current user turn
    messages.append({"role": "user", "content": user_input})

    try:
        # The pipeline call is simpler. We pass the image as the main argument
        # and the text conversation as the `prompt`.
        if current_image:
            # The image goes first, the prompt kwarg contains the conversation history
            output = pipe(current_image, prompt=messages, generate_kwargs={"max_new_tokens": 512, "do_sample": True, "temperature": 0.7})
        else:
            # If no image, the pipeline can work with just the prompt
            output = pipe(prompt=messages, generate_kwargs={"max_new_tokens": 512, "do_sample": True, "temperature": 0.7})
        
        # The pipeline output structure contains the full conversation.
        # We want the content of the last message, which is the model's reply.
        clean_response = output[0]["generated_text"][-1]['content']

    except Exception as e:
        print(f"Caught a critical exception during generation: {e}", flush=True)
        clean_response = (
            "An error occurred during generation. Details:\n\n"
            f"```\n{type(e).__name__}: {e}\n```"
        )

    history_for_display.append((user_input, clean_response))
    return history_for_display, current_image, None, ""

# --- Gradio Interface ---
with gr.Blocks(theme=gr.themes.Soft(), css="footer {display: none !important}") as demo:
    gr.Markdown(
        """
        # AI Symptom Checker powered by MedGemma
        Describe your symptoms below. For visual symptoms (e.g., a skin rash), upload an image. The AI will analyze the inputs and ask clarifying questions if needed.
        """
    )
    
    image_state = gr.State(None)
    chatbot = gr.Chatbot(label="Conversation", height=500, bubble_full_width=False, avatar_images=("user.png", "bot.png"))
    chat_history = gr.State([])

    with gr.Row():
        image_box = gr.Image(type="pil", label="Upload Image of Symptom (Optional)")

    with gr.Row():
        text_box = gr.Textbox(label="Describe your symptoms...", placeholder="e.g., I have a rash on my arm that is red and itchy...", scale=4)
        submit_btn = gr.Button("Send", variant="primary", scale=1)

    def clear_all():
        return [], None, None, ""

    clear_btn = gr.Button("Start New Conversation")
    clear_btn.click(fn=clear_all, outputs=[chat_history, image_state, image_box, text_box], queue=False)
    
    def on_submit(user_input, display_history, new_image, persisted_image):
        if not user_input.strip() and not new_image:
            return display_history, persisted_image, None, ""
        return symptom_checker_chat(user_input, display_history, new_image, persisted_image)

    # Event Handlers for submit button and enter key
    submit_btn.click(
        fn=on_submit,
        inputs=[text_box, chat_history, image_box, image_state],
        outputs=[chat_history, image_state, image_box, text_box]
    )
    text_box.submit(
        fn=on_submit,
        inputs=[text_box, chat_history, image_box, image_state],
        outputs=[chat_history, image_state, image_box, text_box]
    )

if __name__ == "__main__":
    demo.launch(debug=True)