import gradio as gr
import torch
from transformers import AutoProcessor, AutoModelForCausalLM
from PIL import Image
import os
import spaces

# --- Configuration ---
hf_token = os.environ.get("HF_TOKEN")
model_id = "google/medgemma-4b-it"

# --- Model Loading ---
if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8:
    dtype = torch.bfloat16
else:
    dtype = torch.float16

model_loaded = False
try:
    processor = AutoProcessor.from_pretrained(model_id, token=hf_token)
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        token=hf_token,
        torch_dtype=dtype,
        device_map="auto",
    )
    model_loaded = True
    print("Model loaded successfully on device:", model.device)
except Exception as e:
    print(f"Error loading model: {e}")

# --- Core Chatbot Function ---
@spaces.GPU
def symptom_checker_chat(user_input, history_state, new_image_upload, image_state):
    """
    Manages the conversational flow, persisting the image across turns.

    Args:
        user_input (str): Text from the user.
        history_state (list): Stateful conversation history.
        new_image_upload (PIL.Image): A new image uploaded in this turn.
        image_state (PIL.Image): The image persisted from a previous turn.

    Returns:
        tuple: Updated values for all relevant Gradio components.
    """
    if not model_loaded:
        history_state.append((user_input, "Error: The model could not be loaded."))
        return history_state, history_state, None, None, ""

    # FIX: Determine which image to use. A new upload takes precedence.
    # This is the key to solving the "image amnesia" problem.
    current_image = new_image_upload if new_image_upload is not None else image_state
    
    # If this is the *first* turn with an image, add the <image> token.
    # Don't add it again if the image is just being carried over in the state.
    if new_image_upload is not None:
        model_input_text = f"<image>\n{user_input}"
    else:
        model_input_text = user_input

    system_prompt = "You are an expert, empathetic AI medical assistant..." # Keep your detailed prompt

    # Construct the full conversation history
    conversation = [{"role": "system", "content": system_prompt}]
    for turn_input, assistant_output in history_state:
        conversation.append({"role": "user", "content": turn_input})
        if assistant_output:
            conversation.append({"role": "assistant", "content": assistant_output})
    conversation.append({"role": "user", "content": model_input_text})
        
    prompt = processor.apply_chat_template(
        conversation, tokenize=False, add_generation_prompt=True
    )

    # Process inputs. Crucially, pass `current_image` to the processor.
    try:
        if current_image:
            inputs = processor(text=prompt, images=current_image, return_tensors="pt").to(model.device, dtype)
        else:
            inputs = processor(text=prompt, return_tensors="pt").to(model.device, dtype)

        outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
        input_token_len = inputs["input_ids"].shape[1]
        generated_tokens = outputs[:, input_token_len:]
        clean_response = processor.decode(generated_tokens[0], skip_special_tokens=True).strip()

    except Exception as e:
        print(f"Error during model generation: {e}")
        clean_response = "An error occurred during generation. Please check the logs."

    # Update the text history with the model-aware input
    history_state.append((model_input_text, clean_response))
    
    # Create the display history without the special tokens
    display_history = [(turn.replace("<image>\n", ""), resp) for turn, resp in history_state]
    
    # Return everything:
    # 1. display_history -> to the chatbot UI
    # 2. history_state -> to the text history state
    # 3. current_image -> to the image state to be persisted
    # 4. None -> to the image upload box to clear it
    # 5. "" -> to the text box to clear it
    return display_history, history_state, 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\n...") # Keep your intro markdown
    
    # FIX: Add a new state component to hold the image across turns
    image_state = gr.State(None)

    chatbot = gr.Chatbot(label="Conversation", height=500, bubble_full_width=False)
    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...", scale=4)
        submit_btn = gr.Button("Send", variant="primary", scale=1)

    # FIX: Update the clear function to also clear the new image_state
    def clear_all():
        return [], [], None, None, ""

    clear_btn = gr.Button("Start New Conversation")
    clear_btn.click(
        fn=clear_all, 
        outputs=[chatbot, chat_history, image_state, image_box, text_box], 
        queue=False
    )
    
    # Combine submit actions into a single function for DRY principle
    def on_submit(user_input, history, new_image, persisted_image):
        return symptom_checker_chat(user_input, history, new_image, persisted_image)

    # FIX: Update inputs and outputs to include the new image_state
    submit_btn.click(
        fn=on_submit,
        inputs=[text_box, chat_history, image_box, image_state],
        outputs=[chatbot, chat_history, image_state, image_box, text_box]
    )
        
    text_box.submit(
        fn=on_submit,
        inputs=[text_box, chat_history, image_box, image_state],
        outputs=[chatbot, chat_history, image_state, image_box, text_box]
    )

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