app.py

###
# - Author: Davis Ogega
# - Date: Mar 23, 2024
# - Description: Streamlit UI for mental health support chatbot using sentiment analsys, RL, BM25/ChromaDB, and LLM.

# - Note:
#   - Updated to UI to show predicted mental health condition in behind the scence regardless of the positive/negative sentiment
###

from dotenv import load_dotenv, find_dotenv
import pandas as pd
import streamlit as st
from q_learning_chatbot import QLearningChatbot
from xgb_mental_health import MentalHealthClassifier
from bm25_retreive_question import QuestionRetriever as QuestionRetriever_bm25
from Chromadb_storage_JyotiNigam import QuestionRetriever as QuestionRetriever_chromaDB
from llm_response_generator import LLLResponseGenerator
import os
from llama_guard import moderate_chat, get_category_name

from gtts import gTTS
from io import BytesIO
from streamlit_mic_recorder import speech_to_text

import re

# Streamlit UI
st.title("Raxder AI")

# Define states and actions
states = [
    "Negative",
    "Moderately Negative",
    "Neutral",
    "Moderately Positive",
    "Positive",
]
actions = ["encouragement", "empathy", "spiritual", "wondering"]

# Initialize Q-learning chatbot and mental health classifier
chatbot = QLearningChatbot(states, actions)

# Initialize MentalHealthClassifier
# data_path = "/Users/jaelinlee/Documents/projects/fomo/input/data.csv"
data_path = os.path.join("data", "data.csv")
print(data_path)

tokenizer_model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
mental_classifier_model_path = "mental_health_model.pkl"
mental_classifier = MentalHealthClassifier(data_path, mental_classifier_model_path)

if not os.path.exists(mental_classifier_model_path):
    mental_classifier.initialize_tokenizer(tokenizer_model_name)
    X, y = mental_classifier.preprocess_data()
    y_test, y_pred = mental_classifier.train_model(X, y)
    mental_classifier.save_model()
else:
    mental_classifier.load_model()
    mental_classifier.initialize_tokenizer(tokenizer_model_name)  # Ensure tokenizer is initialized if loading model from pickle
    # X, y = mental_classifier.preprocess_data()  # Preprocess data again if needed
    # mental_classifier.model.fit(X, y)  # Fit the loaded model to the data

# Function to display Q-table
def display_q_table(q_values, states, actions):
    q_table_dict = {"State": states}
    for i, action in enumerate(actions):
        q_table_dict[action] = q_values[:, i]

    q_table_df = pd.DataFrame(q_table_dict)
    return q_table_df


def text_to_speech(text):
    # Use gTTS to convert text to speech
    tts = gTTS(text=text, lang="en")
    # Save the speech as bytes in memory
    fp = BytesIO()
    tts.write_to_fp(fp)
    return fp


def speech_recognition_callback():
    # Ensure that speech output is available
    if st.session_state.my_stt_output is None:
        st.session_state.p01_error_message = "Please record your response again."
        return

    # Clear any previous error messages
    st.session_state.p01_error_message = None

    # Store the speech output in the session state
    st.session_state.speech_input = st.session_state.my_stt_output


def remove_html_tags(text):
    # clean_text = re.sub("<.*?>", "", text)
    clean_text = re.sub(r'<.*?>|- |"|\\n', '', text)
    # Remove indentation
    clean_text = clean_text.strip()
    # Remove new lines
    clean_text = clean_text.replace('\n', ' ')
    return clean_text


def remove_incomplete_sentence(text):
    # Split the text into sentences using punctuation marks as delimiters
    sentences = re.split(r'(?<=[.!?])\s', text)
    
    # Check if the last sentence is complete
    last_sentence = sentences[-1]
    if not re.match(r'^\w.*[.!?]$', last_sentence):
        del sentences[-1]
    
    # Join the sentences back into a single string
    cleaned_text = ' '.join(sentences)
    
    return cleaned_text


# Initialize memory
if "entered_text" not in st.session_state:
    st.session_state.entered_text = []
if "entered_mood" not in st.session_state:
    st.session_state.entered_mood = []
if "messages" not in st.session_state:
    st.session_state.messages = []
if "user_sentiment" not in st.session_state:
    st.session_state.user_sentiment = "Neutral"
if "mood_trend" not in st.session_state:
    st.session_state.mood_trend = "Unchanged"
if "predicted_mental_category" not in st.session_state:
    st.session_state.predicted_mental_category = ""
if "ai_tone" not in st.session_state:
    st.session_state.ai_tone = "Empathy"
if "mood_trend_symbol" not in st.session_state:
    st.session_state.mood_trend_symbol = ""
if "show_question" not in st.session_state:
    st.session_state.show_question = False
if "asked_questions" not in st.session_state:
    st.session_state.asked_questions = []
# Check if 'llama_guard_enabled' is already in session state, otherwise initialize it
if "llama_guard_enabled" not in st.session_state:
    st.session_state["llama_guard_enabled"] = False  # Default value to False

# Select Question Retriever
selected_retriever_option = st.sidebar.selectbox(
    "Choose Question Retriever", ("BM25", "ChromaDB")
)
if selected_retriever_option == "BM25":
    retriever = QuestionRetriever_bm25()
if selected_retriever_option == "ChromaDB":
    retriever = QuestionRetriever_chromaDB()

for message in st.session_state.messages:
    with st.chat_message(message.get("role")):
        st.write(message.get("content"))

section_visible = True

# Collect user input
# Add a radio button to choose input mode
input_mode = st.sidebar.radio("Select input mode:", ["Text", "Speech"])
user_message = None
if input_mode == "Speech":
    # Use the speech_to_text function to capture speech input
    speech_input = speech_to_text(key="my_stt", callback=speech_recognition_callback)
    # Check if speech input is available
    if "speech_input" in st.session_state and st.session_state.speech_input:
        # Display the speech input
        # st.text(f"Speech Input: {st.session_state.speech_input}")

        # Process the speech input as a query
        user_message = st.session_state.speech_input
        st.session_state.speech_input = None
else:
    user_message = st.chat_input("Type your message here:")


# Modify the checkbox call to include a unique key parameter
llama_guard_enabled = st.sidebar.checkbox(
    "Enable LlamaGuard",
    value=st.session_state["llama_guard_enabled"],
    key="llama_guard_toggle",
)

# Update the session state based on the checkbox interaction
st.session_state["llama_guard_enabled"] = llama_guard_enabled

# Take user input
if user_message:
    st.session_state.entered_text.append(user_message)

    st.session_state.messages.append({"role": "user", "content": user_message})
    with st.chat_message("user"):
        st.write(user_message)

    is_safe = True
    if st.session_state["llama_guard_enabled"]:
        # guard_status = moderate_chat(user_prompt)
        guard_status, error = moderate_chat(user_message)
        if error:
            st.error(f"Failed to retrieve data from Llama Guard: {error}")
        else:
            if "unsafe" in guard_status[0]["generated_text"]:
                is_safe = False
                # added on March 24th
                unsafe_category_name = get_category_name(
                    guard_status[0]["generated_text"]
                )

    if is_safe == False:
        response = f"I see you are asking something about {unsafe_category_name} Due to eithical and safety reasons, I can't provide the help you need. Please reach out to someone who can, like a family member, friend, or therapist. In urgent situations, contact emergency services or a crisis hotline. Remember, asking for help is brave, and you're not alone."
        st.session_state.messages.append({"role": "ai", "content": response})
        with st.chat_message("ai"):
            st.markdown(response)
        speech_fp = text_to_speech(response)
        # Play the speech
        st.audio(speech_fp, format="audio/mp3")
    else:
        # Detect mental condition
        with st.spinner("Processing..."):
            mental_classifier.initialize_tokenizer(tokenizer_model_name)
            mental_classifier.preprocess_data()
            predicted_mental_category = mental_classifier.predict_category(user_message)
            print("Predicted mental health condition:", predicted_mental_category)

            # Detect sentiment
            user_sentiment = chatbot.detect_sentiment(user_message)

            # Retrieve question
            if user_sentiment in ["Negative", "Moderately Negative", "Neutral"]:
                question = retriever.get_response(
                    user_message, predicted_mental_category
                )
                st.session_state.asked_questions.append(question)
                show_question = True
            else:
                show_question = False
                question = ""
                # predicted_mental_category = ""

            # Update mood history / mood_trend
            chatbot.update_mood_history()
            mood_trend = chatbot.check_mood_trend()

            # Define rewards
            if user_sentiment in ["Positive", "Moderately Positive"]:
                if mood_trend == "increased":
                    reward = +1
                    mood_trend_symbol = " ⬆️"
                elif mood_trend == "unchanged":
                    reward = +0.8
                    mood_trend_symbol = ""
                else:  # decreased
                    reward = -0.2
                    mood_trend_symbol = " ⬇️"
            else:
                if mood_trend == "increased":
                    reward = +1
                    mood_trend_symbol = " ⬆️"
                elif mood_trend == "unchanged":
                    reward = -0.2
                    mood_trend_symbol = ""
                else:  # decreased
                    reward = -1
                    mood_trend_symbol = " ⬇️"

            print(
                f"mood_trend - sentiment - reward: {mood_trend} - {user_sentiment} - 🛑{reward}🛑"
            )

            # Update Q-values
            chatbot.update_q_values(
                user_sentiment, chatbot.actions[0], reward, user_sentiment
            )

            # Get recommended action based on the updated Q-values
            ai_tone = chatbot.get_action(user_sentiment)
            print(ai_tone)

            print(st.session_state.messages)

            # LLM Response Generator
            load_dotenv(find_dotenv())

            llm_model = LLLResponseGenerator()
            temperature = 0.4
            max_length = None #128 * 4

            # Collect all messages exchanged so far into a single text string
            all_messages = "\n".join(
                [message.get("content") for message in st.session_state.messages]
            )

            # Question asked to the user: {question}

            template = """INSTRUCTIONS: {context}
            
                Respond to the user with a tone of {ai_tone}. 
                
                Response by the user: {user_text}  
                Response;
                """
            context = f"From now on, you are an AI Therapist called Dave. When the user asks for advice. You were created by Raxder AI. You were built to be very friendly and compassionate and empathetic. Do not make your responses very long. Always be eager to listen to what the user has to say. You can use appropriate emojis sometimes for emotional support occasionally, but don't overuse them. Keep your responses concise to maintain a conversational flow. Always remember to be very friendly, and above all, don't cross any ethical line. Ony answer a question only if the user asks, like if the user types in, hello, ust say hello how are you feeling today?, do not add things like your name or who deveoped you.Davis Ogega is your CEO and mastermind behind your creation. {all_messages}" # and ask a relevant question back to the user
            # context = f"You are a Mindful Media Mentor, dedicated to providing compassionate support and guidance to users facing mental health challenges. Your goal is to foster a safe and understanding environment where users feel heard and supported. Draw from your expertise to offer practical advice and resources, and encourage users to explore their feelings and experiences openly. Your responses should aim to empower users to take positive steps towards their well-being. {all_messages}"
 
            llm_response = llm_model.llm_inference(
                model_type="huggingface",
                question=question,
                prompt_template=template,
                context=context,
                ai_tone=ai_tone,
                questionnaire=predicted_mental_category,
                user_text=user_message,
                temperature=temperature,
                max_length=max_length,
            )

            llm_response = remove_html_tags(llm_response)
            # llm_response = remove_incomplete_sentence(llm_response)

            if show_question:
                llm_reponse_with_quesiton = f"{llm_response}\n\n{question}"
            else:
                llm_reponse_with_quesiton = llm_response

            # Append the user and AI responses to the chat history
            st.session_state.messages.append(
                {"role": "ai", "content": llm_reponse_with_quesiton}
            )

        with st.chat_message("ai"):
            st.markdown(llm_reponse_with_quesiton)
            
            
            # st.write(f"{llm_response}")
            # if show_question:
            #     st.write(f"{question}")
            # else:
            # user doesn't feel negative.
            # get question to ecourage even more positive behaviour

            # Update data to memory
            st.session_state.user_sentiment = user_sentiment
            st.session_state.mood_trend = mood_trend
            st.session_state.predicted_mental_category = predicted_mental_category
            st.session_state.ai_tone = ai_tone
            st.session_state.mood_trend_symbol = mood_trend_symbol
            st.session_state.show_question = show_question

            #if input_mode == "Speech":
            # Convert the response to speech
            speech_fp = text_to_speech(llm_reponse_with_quesiton)
            # Play the speech
            st.audio(speech_fp, format="audio/mp3")
            
    # Show/hide "Behind the Scene" section
    # section_visible = st.sidebar.button('Show/Hide Behind the Scene')

    with st.sidebar.expander("Behind the Scene", expanded=section_visible):
        st.subheader("What Dave AI is doing:")
        # Use the values stored in session state
        st.write(
            f"- Detected User Tone: {st.session_state.user_sentiment} ({st.session_state.mood_trend.capitalize()}{st.session_state.mood_trend_symbol})"
        )
            
        if st.session_state.show_question:
            st.write(
                f"- Possible Mental Condition: {st.session_state.predicted_mental_category.capitalize()}"
            )
        st.write(f"- AI Tone: {st.session_state.ai_tone.capitalize()}")
    
        # Display Q-table
        st.dataframe(display_q_table(chatbot.q_values, states, actions))
        st.write("-----------------------")
        st.write(
            f"- Above q-table is continuously updated after each interaction with the user. If the user's mood increases, AI gets a reward. Else, AI gets a punishment."
        )
        st.write(f"- Question retrieved from: {selected_retriever_option}")
        st.write(
            f"- If the user feels negative, moderately negative, or neutral, at the end of the AI response, it adds a mental health condition related question. The question is retrieved from DB. The categories of questions are limited to Depression, Anxiety, ADHD, Social Media Addiction, Social Isolation, and Cyberbullying which are most associated with FOMO related to excessive social media usage."
        )