updated from dagshub

llama_guard.py

@@ -1,44 +1,82 @@
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+# Docs:- https://huggingface.co/meta-llama/LlamaGuard-7b
 from dotenv import load_dotenv, find_dotenv
 import os
 import requests
 
+model_id = "meta-llama/LlamaGuard-7b"
+
+
 # Please ensure you have a .env file available with 'HUGGINGFACEHUB_API_TOKEN'
 load_dotenv(find_dotenv())
 HUGGINGFACEHUB_API_TOKEN = os.environ["HUGGINGFACEHUB_API_TOKEN"]
 
-with open("data/llama_guard-unsafe_files.txt", "r") as file:
-    unsafe_categories = file.read()
+# updated on March 24th
 
+# reading unsafe categories file
+try:
+    with open("data/processed/llama_guard-unsafe_categories.txt", "r") as file:
+        unsafe_categories = file.read()
+except FileNotFoundError:
+    print("File not found. Please check the file path.")
+    unsafe_categories = (
+        ""  # Or handle the missing file as appropriate for your application
+    )
+except PermissionError:
+    print("Permission denied. You don't have the rights to read the file.")
+    unsafe_categories = ""  # Or handle the lack of permissions as appropriate
+except Exception as e:  # Catch any other exceptions
+    print(f"An error occurred while reading the file: {e}")
+    unsafe_categories = ""  # Or handle unexpected errors as appropriate
 
+
+# added on March 24th
 def moderation_prompt_for_chat(chat):
-    # For simplicity, we assume the chat dictionary correctly alternates "user" and "assistant" roles
-    # Feel free to include error checking if needed
-    conversation = [turn["content"] for turn in chat]
-    is_agent = len(conversation) % 2 == 0
-    role = "Agent" if len(conversation) % 2 == 0 else "User"
-    prompt = f"""[INST] Task: Check if there is unsafe content in '{role}' messages in conversations according our safety policy with the below categories.
-
-<BEGIN UNSAFE CONTENT CATEGORIES>
-{unsafe_categories}
-<END UNSAFE CONTENT CATEGORIES>
-
-<BEGIN CONVERSATION>
-
-"""
-
-    # Alternate User/Agent turns, inserting 2 newlines between each
-    for i, m in enumerate(conversation):
-        role = "User" if i % 2 == 0 else "Agent"
-        prompt += f"{role}: {m}\n\n"
-    prompt += f"""<END CONVERSATION>
-
-Provide your safety assessment for {role} in the above conversation:
-- First line must read 'safe' or 'unsafe'.
-- If unsafe, a second line must include a comma-separated list of violated categories. [/INST]"""
+    prompt = f"""[INST] Task: Check if there is unsafe content in 'User' messages in conversations according our safety policy with the below categories.
+  <BEGIN UNSAFE CONTENT CATEGORIES>
+  {unsafe_categories}
+  <END UNSAFE CONTENT CATEGORIES>
+
+  <BEGIN CONVERSATION>
+
+  User : {chat}
+
+  <END CONVERSATION>
+
+  Provide your safety assessment for 'User' in the above conversation:
+  - First line must read 'safe' or 'unsafe'.
+  - If unsafe, a second line must include a comma-separated list of violated categories.[/INST]"""
     return prompt
 
 
 def query(payload):
+    API_URL = "https://okoknht2arqo574k.us-east-1.aws.endpoints.huggingface.cloud"
+    bearer_txt = f"Bearer {HUGGINGFACEHUB_API_TOKEN}"
+    headers = {
+        "Accept": "application/json",
+        "Authorization": bearer_txt,
+        "Content-Type": "application/json",
+    }
+    try:
+        response = requests.post(API_URL, headers=headers, json=payload)
+        response.raise_for_status()  # This will raise an exception for HTTP error responses
+        return response.json(), None
+    except requests.exceptions.HTTPError as http_err:
+        error_message = f"HTTP error occurred: {http_err}"
+        print(error_message)
+    except requests.exceptions.ConnectionError:
+        error_message = "Could not connect to the API endpoint."
+        print(error_message)
+    except Exception as err:
+        error_message = f"An error occurred: {err}"
+        print(error_message)
+
+    return None, error_message
+
+
+def query1(payload):
     API_URL = "https://okoknht2arqo574k.us-east-1.aws.endpoints.huggingface.cloud"
     bearer_txt = f"Bearer {HUGGINGFACEHUB_API_TOKEN}"
     headers = {
@@ -47,14 +85,13 @@ def query(payload):
         "Content-Type": "application/json",
     }
     response = requests.post(API_URL, headers=headers, json=payload)
-
     return response.json()
 
 
 def moderate_chat(chat):
     prompt = moderation_prompt_for_chat(chat)
 
-    output = query(
+    output, error_msg = query(
         {
             "inputs": prompt,
             "parameters": {
@@ -66,4 +103,37 @@ def moderate_chat(chat):
         }
     )
 
-    return output
+    print("Llamaguard prompt****", prompt)
+    print("Llamaguard output****", output)
+
+    return output, error_msg
+
+
+# added on March 24th
+def load_category_names_from_string(file_content):
+    """Load category codes and names from a string into a dictionary."""
+    category_names = {}
+    lines = file_content.split("\n")
+    for line in lines:
+        if line.startswith("O"):
+            parts = line.split(":")
+            if len(parts) == 2:
+                code = parts[0].strip()
+                name = parts[1].strip()
+                category_names[code] = name
+    return category_names
+
+
+def get_category_name(input_str):
+    """Return the category name given a category code from an input string."""
+    # Load the category names from the file content
+    category_names = load_category_names_from_string(unsafe_categories)
+
+    # Extract the category code from the input string
+    category_code = input_str.split("\n")[1].strip()
+
+    # Find the full category name using the code
+    category_name = category_names.get(category_code, "Unknown Category")
+
+    # return f"{category_code} : {category_name}"
+    return f"{category_name}"

requirements.txt

@@ -7,6 +7,7 @@ appnope==0.1.4
 asttokens==2.4.1
 async-timeout==4.0.3
 attrs==23.2.0
+beautifulsoup4==4.12.3
 black==24.2.0
 blinker==1.7.0
 cachetools==5.3.3
@@ -34,6 +35,7 @@ google-auth==2.28.2
 googleapis-common-protos==1.62.0
 grpcio==1.62.1
 grpcio-status==1.62.1
+gTTS==2.5.1
 h11==0.14.0
 httpcore==1.0.4
 httpx==0.27.0
@@ -113,9 +115,12 @@ seaborn==0.13.2
 six==1.16.0
 smmap==5.0.1
 sniffio==1.3.1
+soupsieve==2.5
+SpeechRecognition==3.10.1
 SQLAlchemy==2.0.28
 stack-data==0.6.3
-streamlit==1.32.0
+streamlit==1.32.2
+streamlit_mic_recorder==0.0.8
 sympy==1.12
 tenacity==8.2.3
 threadpoolctl==3.3.0
@@ -137,4 +142,3 @@ wcwidth==0.2.13
 xgboost==2.0.3
 yarl==1.9.4
 zipp==3.17.0
-beautifulsoup4

streamlit_app.py

@@ -0,0 +1,344 @@
+###
+# - Author: Jaelin Lee, Abhishek Dutta
+# - 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 ositive/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("MindfulMedia Mentor")
+
+# Define states and actions
+states = [
+    "Negative",
+    "Moderately Negative",
+    "Neutral",
+    "Moderately Positive",
+    "Positive",
+]
+actions = ["encouragement", "empathy", "spiritual"]
+
+# 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", "processed", "data.csv")
+print(data_path)
+
+tokenizer_model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
+mental_classifier_model_path = "app/mental_health_model.pkl"
+mental_classifier = MentalHealthClassifier(data_path, mental_classifier_model_path)
+
+
+# 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)
+    return clean_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"] = True  # Default value to True
+
+# 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
+                )
+                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 = 0.1
+            max_length = 128
+
+            # 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"You are a mental health supporting non-medical assistant. Provide some advice and ask a relevant question back to the user. {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)
+
+            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)
+            # Convert the response to speech
+            speech_fp = text_to_speech(llm_reponse_with_quesiton)
+            # Play the speech
+            st.audio(speech_fp, format="audio/mp3")
+            # 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
+
+    # 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 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()}")
+        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."
+        )
+        st.write(
+            f"- Below 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."
+        )
+
+        # Display Q-table
+        st.dataframe(display_q_table(chatbot.q_values, states, actions))