app.py

import openai
import gradio as gr
from gradio.components import Audio, Textbox
import os
import re
import tiktoken
from transformers import GPT2Tokenizer
import whisper
import pandas as pd
from datetime import datetime, timezone, timedelta
import notion_df
import concurrent.futures
import nltk
from nltk.tokenize import sent_tokenize
nltk.download('punkt')
import spacy
from spacy import displacy
from gradio import Markdown
import threading

# Define the tokenizer and model
tokenizer = GPT2Tokenizer.from_pretrained('gpt2-medium')
model = openai.api_key = os.environ["OPENAI_API_KEY"]

# Define the initial message and messages list
initialt = 'You are a USMLE Tutor. Respond with ALWAYS layered "bullet points" (listing rather than sentences) to all input with a fun mneumonics to memorize that list. But you can answer up to 1200 words if the user requests longer response.'
initial_message = {"role": "system", "content": initialt}
messages = [initial_message]
messages_rev = [initial_message]

# Define the answer counter
answer_count = 0

# Define the Notion API key
API_KEY = os.environ["API_KEY"]


nlp = spacy.load("en_core_web_sm")

def process_nlp(system_message):
    # Colorize the system message text
    colorized_text = colorize_text(system_message['content'])
    return colorized_text

from colour import Color

# define color combinations for different parts of speech
COLORS = {
    "NOUN": "#000000",  # Black
    "VERB": "#ff6936",  # Orange
    "ADJ": "#4363d8",   # Blue
    "ADV": "#228b22",   # Green
    "digit": "#9a45d6", # Purple
    "punct": "#ffcc00", # Yellow
    "quote": "#b300b3"  # Magenta
}

# define color combinations for individuals with dyslexia and color vision deficiencies
DYSLEXIA_COLORS = {
    "NOUN": "#000000",
    "VERB": "#ff6936",
    "ADJ": "#4363d8",
    "ADV": "#228b22",
    "digit": "#9a45d6",
    "punct": "#ffcc00",
    "quote": "#b300b3",
}
RED_GREEN_COLORS = {
    "NOUN": "#000000",
    "VERB": "#fe642e",  # Lighter orange
    "ADJ": "#2e86c1",   # Lighter blue
    "ADV": "#82e0aa",   # Lighter green
    "digit": "#aa6c39", # Brown
    "punct": "#f0b27a", # Lighter yellow
    "quote": "#9932cc"  # Darker magenta
}

# define a muted background color
BACKGROUND_COLOR = "#ffffff"  # White

# define font and size
FONT = "OpenDyslexic"
FONT_SIZE = "18px"

def colorize_text(text, colors=DYSLEXIA_COLORS, background_color=None, font=FONT, font_size=FONT_SIZE):
    if colors is None:
        colors = COLORS
    colorized_text = ""
    lines = text.split("\n")
    
    # set background color
    if background_color is None:
        background_color = BACKGROUND_COLOR
    
    # iterate over the lines in the text
    for line in lines:
        # parse the line with the language model
        doc = nlp(line)
        # iterate over the tokens in the line
        for token in doc:
            # check if the token is an entity
            if token.ent_type_:
                # use dyslexia colors for entity if available
                if colors == COLORS:
                    color = DYSLEXIA_COLORS.get(token.pos_, None)
                else:
                    color = colors.get(token.pos_, None)
                # check if a color is available for the token
                if color is not None:
                    colorized_text += (
                        f'<span style="color: {color}; '
                        f'background-color: {background_color}; '
                        f'font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'font-weight: bold; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
                else:
                    colorized_text += (
                        f'<span style="font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'font-weight: bold; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
            else:
                # check if a color is available for the token
                color = colors.get(token.pos_, None)
                if color is not None:
                    colorized_text += (
                        f'<span style="color: {color}; '
                        f'background-color: {background_color}; '
                        f'font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'font-weight: bold; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
                elif token.is_digit:
                    colorized_text += (
                        f'<span style="color: {colors["digit"]}; '
                        f'background-color: {background_color}; '
                        f'font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'font-weight: bold; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
                elif token.is_punct:
                    colorized_text += (
                        f'<span style="color: {colors["punct"]}; '
                        f'background-color: {background_color}; '
                        f'font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'font-weight: bold; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
                elif token.is_quote:
                    colorized_text += (
                        f'<span style="color: {colors["quote"]}; '
                        f'background-color: {background_color}; '
                        f'font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
                else:
                    # use larger font size for specific parts of speech, such as nouns and verbs
                    font_size = FONT_SIZE
                    if token.pos_ in ["NOUN", "VERB"]:
                        font_size = "22px"
                    colorized_text += (
                        f'<span style="font-family: {font}; '
                        f'font-size: {font_size}; '
                        f'font-weight: bold; '
                        f'text-decoration: none; '
                        f'padding-right: 0.5em;">'
                        f"{token.text}</span>"
                    )
        colorized_text += "<br>"
    return colorized_text


def colorize_and_update(system_message, submit_update):
    colorized_system_message = colorize_text(system_message['content'])
    submit_update(None, colorized_system_message)  # Pass the colorized_system_message as the second output

def update_text_output(system_message, submit_update):
    submit_update(system_message['content'], None)
    
def train(text):
    now_et = datetime.now(timezone(timedelta(hours=-4)))
    published_date = now_et.strftime('%m-%d-%y %H:%M')
    df = pd.DataFrame([text])
    notion_df.upload(df, 'https://www.notion.so/US-62e861a0b35f43da8ef9a7789512b8c2?pvs=4', title=str(published_date), api_key=API_KEY)


def transcribe(audio, text, submit_update=None):
    global messages
    global answer_count
    transcript = {'text': ''} 
    input_text = []
    
    # Check if the first word of the first line is "COLORIZE"
    if text and text.split("\n")[0].split(" ")[0].strip().upper() == "COLORIZE":
        train(text)
        colorized_input = colorize_text(text)
        return text, colorized_input
    
    # Transcribe the audio if provided
    if audio is not None:
        audio_file = open(audio, "rb")
        transcript = openai.Audio.transcribe("whisper-1", audio_file, language="en")
        
    # Tokenize the text input
    if text is not None:
        # Split the input text into sentences
        sentences = re.split("(?<=[.!?]) +", text)
    
        # Initialize a list to store the tokens
        input_tokens = []
    
        # Add each sentence to the input_tokens list
        for sentence in sentences:
            # Tokenize the sentence using the GPT-2 tokenizer
            sentence_tokens = tokenizer.encode(sentence)
            # Check if adding the sentence would exceed the token limit
            if len(input_tokens) + len(sentence_tokens) < 1440:
                # Add the sentence tokens to the input_tokens list
                input_tokens.extend(sentence_tokens)
            else:
                # If adding the sentence would exceed the token limit, truncate it
                sentence_tokens = sentence_tokens[:1440-len(input_tokens)]
                input_tokens.extend(sentence_tokens)
                break
        # Decode the input tokens into text
        input_text = tokenizer.decode(input_tokens)
    
    # Add the input text to the messages list
    messages.append({"role": "user", "content": transcript["text"]+input_text})

    # Check if the accumulated tokens have exceeded 2096
    num_tokens = sum(len(tokenizer.encode(message["content"])) for message in messages)
    if num_tokens > 2096:
        # Concatenate the chat history
        chat_transcript = "\n\n".join([f"[ANSWER {answer_count}]{message['role']}: {message['content']}" for message in messages if message['role'] != 'system'])

        # Append the number of tokens used to the end of the chat transcript
        chat_transcript += f"\n\nNumber of tokens used: {num_tokens}\n\n"

        # Get the current time in Eastern Time (ET)
        now_et = datetime.now(timezone(timedelta(hours=-4)))
        # Format the time as string (YY-MM-DD HH:MM)
        published_date = now_et.strftime('%m-%d-%y %H:%M')

        # Upload the chat transcript to Notion
        df = pd.DataFrame([chat_transcript])
        notion_df.upload(df, 'https://www.notion.so/US-62e861a0b35f43da8ef9a7789512b8c2?pvs=4', title=str(published_date+'FULL'), api_key=API_KEY)

        messages = [initial_message]
        messages.append({"role": "user", "content": initialt})
        answer_count = 0
        # Add the input text to the messages list
        messages.append({"role": "user", "content": input_text})
    else:
        # Increment the answer counter
        answer_count += 1

    # Generate the system message using the OpenAI API
    with concurrent.futures.ThreadPoolExecutor() as executor:
        prompt = [{"text": f"{message['role']}: {message['content']}\n\n"} for message in messages]
        system_message = openai.ChatCompletion.create(
            model="gpt-3.5-turbo",
            messages=messages,
            max_tokens=2000
        )["choices"][0]["message"]
    # Wait for the completion of the OpenAI API call
        
    if submit_update:  # Check if submit_update is not None
        update_text_output(system_message, submit_update)

    # Add the system message to the messages list
    messages.append(system_message)

    # Add the system message to the beginning of the messages list
    messages_rev.insert(0, system_message)
    # Add the input text to the messages list
    messages_rev.insert(0, {"role": "user", "content": input_text + transcript["text"]})

    # Start a separate thread to process the colorization and update the Gradio interface
    if submit_update:  # Check if submit_update is not None
        colorize_thread = threading.Thread(target=colorize_and_update, args=(system_message, submit_update))
        colorize_thread.start()

    # Concatenate the chat history
    chat_transcript = "\n\n".join([f"[ANSWER {answer_count}]{message['role']}: {message['content']}" for message in messages_rev if message['role'] != 'system'])
    
    # Append the number of tokens used to the end of the chat transcript
    chat_transcript += f"\n\nNumber of tokens used: {num_tokens}\n\n"
    
    # Save the chat transcript to a file
    with open("conversation_history.txt", "a") as f:
        f.write(chat_transcript)
    
    # Upload the chat transcript to Notion
    now_et = datetime.now(timezone(timedelta(hours=-4)))
    published_date = now_et.strftime('%m-%d-%y %H:%M')
    df = pd.DataFrame([chat_transcript])
    notion_df.upload(df, 'https://www.notion.so/US-62e861a0b35f43da8ef9a7789512b8c2?pvs=4', title=str(published_date), api_key=API_KEY)
    
    # Return the chat transcript    
    return system_message['content'], colorize_text(system_message['content'])

    
# Define the input and output components for Gradio
audio_input = Audio(source="microphone", type="filepath", label="Record your message")
text_input = Textbox(label="Type your message", max_length=4096)
output_text = Textbox(label="Text Output")
output_html = Markdown()
output_audio = Audio()

# Define the Gradio interface
iface = gr.Interface(
    fn=transcribe,
    inputs=[audio_input, text_input],
    outputs=[output_text, output_html],
    title="Hold On, Pain Ends (HOPE)",
    description="Talk to Your USMLE Tutor HOPE. \n If you want to colorize your note, type COLORIZE in the first line of your input.",
    theme="compact",
    layout="vertical",
    allow_flagging=False
    )

# Run the Gradio interface
iface.launch()