index.py

import openai
import streamlit as st
from annotated_text import annotated_text
import time
import json
import random
from youtube_transcript_api import YouTubeTranscriptApi
import random

import keys

### State example
# TODO: 
# A. Move options selection/enablement state into questions key. 
# B. Remove nested questions.questions
# C. After the answered question (irrespective of if correct or wrong), also allow user to mark the option as invalid

example_state = {
  "render_questions_btn": False,
  "questions": {
    "questions": [
      {
        "question": "What is prompt engineering?",
        "options": [
          "The art of writing good intentional prompts that produce an output from a generative AI model",
          "A more abstract version of programming",
          "A super abstract programming of an AI model",
          "A way to modify the mode or type of task that has been formed"
        ],
        "reference": "",
        "correct_option_index": 0,
        "was_answered": False,
        "was_skipped": False
      },
      {
        "question": "What is the best way to think about prompt engineering?",
        "options": [
          "As a way to modify the mode or type of task that has been formed",
          "As a more abstract version of programming",
          "As a super abstract programming of an AI model",
          "As the next step in programming languages"
        ],
        "reference": "",
        "correct_option_index": 1,
        "was_answered": False,
        "was_skipped": False
      }
    ],
    "questionnaire_state": {
      "questions_asked": 0,
      "questions_skipped": 0
    }
  },
  "options_state": {
    "q1": {
      "1": {
        "selected": False,
        "disabled": True
      },
      "2": {
        "selected": True,
        "disabled": True
      },
      "3": {
        "selected": False,
        "disabled": True
      },
      "4": {
        "selected": False,
        "disabled": True
      },
      "skip_q1": {
        "selected": False,
        "disabled": True
      }
    },
    "q2": {
      "1": {
        "selected": True,
        "disabled": True
      },
      "2": {
        "selected": False,
        "disabled": True
      },
      "3": {
        "selected": False,
        "disabled": True
      },
      "4": {
        "selected": False,
        "disabled": True
      },
      "skip_q2": {
        "selected": False,
        "disabled": True
      }
    }
  },
  "generate_q_btn": {
    "disabled": True
  },
  "user_score": 0,
  "questions_rendered": True
}

openai.organization = keys.OPENAI_ORG
openai.api_key = keys.OPENAI_KEY

### Constants
PROJECT_NAME = 'Memorate' # alt.: Memembered etc
PROJECT_IDEA = 'Improve memorization by answering questions'
TEXT_AREA_INSTRUCTION = 'Paste some text and get asked about it. Max 1300 characters'
# https://www.ukrinform.net/rubric-society/3671800-ukraine-removed-from-list-of-countries-where-chat-gpt-is-not-available.html
DUMMY_TEXT = 'Ukraine has been removed from the list of countries where the artificial intelligence chatbot ChatGPT is blocked. Ukrainian Deputy Prime Minister - Minister of Digital Transformation Mykhailo Fedorov announced this on Telegram, Ukrinform reports. "ChatGPT is now available in Ukraine. The team of the Ministry of Digital Transformation worked for a long time on this decision - official letters, calls and a meeting with the management. Finally, we managed to correct the injustice. Ukraine was removed from the list of countries where ChatGPT is blocked," he wrote. The program will not work only in the territories temporarily occupied by Russia. ChatGPT is an artificial intelligence chatbot developed by AI startup OpenAI and capable of working in dialog mode, which supports requests in different languages.'
MAX_INPUT_LEN_CHARS = 1300
SEPARATOR = '<highlight>'
CORRECT_MARK = '^'
OVERLAP_CHARS = 50 # how many characters before and after the segment should be taken

### Functions
def get_openai_completion(prompt):
    print('\n\nOPENAI-PROMPT:', prompt)
    try:
        res = openai.Completion.create(engine="text-davinci-003", prompt=prompt, max_tokens=400, temperature=0.1)
        print('\n\nOPENAI-RESPONSE:', res)
        return res.choices[0]['text'] or None
    except Exception as e:
        print(f'get_openai_completion - exception: ', e)
        st.write(e)

@st.cache_data
def openai_get_proofs(answer_options, input_text):
    options_prompt = []
    for idx, option in enumerate(answer_options, start=1):
        options_prompt.append(f'{idx}. {option}\n')
    prompt = f"""In the following text find the sentences which correspond to the following phrases:\n{''.join(options_prompt)}\n{input_text.rstrip('.')}
    """
    # OpenAI response for the text input
    # return "\n\nQ1. What is ChatGPT?\nA. An artificial intelligence chatbot ^\nB. A program developed by AI startup OpenAI\nC. A chatbot capable of working in dialog mode\n\nQ2. What did the team of the Ministry of Digital Transformation do to make ChatGPT available in Ukraine?\nA. Sent official letters ^\nB. Held a meeting with the management\nC. Called the management\n\nQ3. Where will ChatGPT not work?\nA. In Ukraine\nB. In the territories temporarily occupied by Russia ^\nC. In different languages"
    
    # OpenAI response for Youtube video captions
    # return "\n1. \"We're also going to be able to use something called a generator model to generate a human natural language answer to our question based on these documents that we've retrieved from an external source.\"\n2. \"Abstractive\"\n3. \"We're going to take all of this and we're going to encode it using what's called a retriever model.\""
    return get_openai_completion(prompt)

@st.cache_data
def openai_get_questions(input_text, questions_number=3, response_options_number=3):
    prompt_v1 = f'''Ask {questions_number} multi-choice questions to the following text and mark the correct answer with "{CORRECT_MARK}":\n\n{input_text}'''
    prompt_v2 = f'''Ask {questions_number} multi-choice questions with {response_options_number} answer options each to the following text. Mark the correct answer with "{CORRECT_MARK}"
    
    {input_text}
    '''
    prompt_v3 = f'''Ask {questions_number} multi-choice questions to the following text. Each question should have {response_options_number} answer options. Mark the correct answer with "{CORRECT_MARK}" at the end:
    
    {input_text}
    '''
    # OpenAI response for the text input
    # return "\n1. ChatGPT is an artificial intelligence chatbot developed by AI startup OpenAI and capable of working in dialog mode, which supports requests in different languages\n2. Ukrainian Deputy Prime Minister - Minister of Digital Transformation Mykhailo Fedorov announced this on Telegram, Ukrinform reports. \"ChatGPT is now available in Ukraine. The team of the Ministry of Digital Transformation worked for a long time on this decision - official letters, calls and a meeting with the management.\n3. The program will not work only in the territories temporarily occupied by Russia."
    
    # OpenAI response for Youtube video captions
    # return "\nQ1. What are we going to be able to do with the components we are using?\nA. Generate a human natural language answer to our question^\nB. Retrieve documents from an external source\nC. Create a GPT model to answer questions\n\nQ2. What type of question and answering are we discussing?\nA. Abstractive^\nB. Generative\nC. Natural language\n\nQ3. What type of model are we using to encode our documents?\nA. GPT model\nB. Retriever model^\nC. Vector model"
    return get_openai_completion(prompt_v3)

# removes the leading "Q1: " for questions,
# or "A. " for response options,
# or "1. " for proofs; might not be 100% reliable
def remove_string_identifier(raw_q):
    parts = raw_q.split(' ')
    without_q_number = parts[1:]
    return ' '.join(without_q_number)

# in some queries to OpenAI all multi-choice questions
# generated had correct answer as the 1st option - not good
def shuffle_options(options, old_correct_idx):
    correct_option = options[old_correct_idx]
    randomised_options = random.sample(options, len(options))
    new_correct_idx = randomised_options.index(correct_option)
    return randomised_options, new_correct_idx

# Expects a string like
# "\n\nQ1: The 1st question?\nA. Option A, marked as correct with^\nB. The 2nd answer\nC. There may be 3-4 options\n\nQ2: Question #2\nA. ...."
# Returns an array of question objects (with keys questions, options, reference, correct_option_index, was_answered, was_skipped)
def openai_res_to_questions(openai_resp_questions):
    res = []
    leading_trailing_new_lines_removed = openai_resp_questions.strip('\n')
    paragraphs = leading_trailing_new_lines_removed.split('\n\n')
    
    for paragraph in paragraphs:
        if not paragraph: continue
        parts = paragraph.split('\n')
        the_question = remove_string_identifier(parts[0]) # remove leading "Q1: " or "1. "
        options_with_numbers = parts[1:]
        options = [remove_string_identifier(option) for option in options_with_numbers] # remove eg "A. "
        options_randomised = random.sample(options, len(options)) # sometimes all the questions may have correct answer at idx 0
        q = {}
        for idx, el in enumerate(options_randomised):
            if CORRECT_MARK in el:
                q['correct_option_index'] = idx
                options_randomised[idx] = el.replace(CORRECT_MARK, '').rstrip()
        q['question'] = the_question
        q['options'] = options_randomised
        q['was_answered'] = False
        q['was_skipped'] = False
        res.append(q)
    return res

# Expects a string like
# "\n\n1. \"Some response, in double quotes or without them\"\n2. \"The 2nd proof\"
# The number of proofs should correspond to the number of questions
# Proofs may not 100% coincide with the original text. My observation: usually the ending
# of the returned proof coincides Ok, while the beginning may be rephrased
# Returns a list of (segments of the original text with the proofs + some overlapping text)
# May consider also using non-OpenAI approaches like
# https://stackoverflow.com/questions/17740833/checking-fuzzy-approximate-substring-existing-in-a-longer-string-in-python
def openai_res_to_references(openai_resp_proofs, input_text):
    paragraphs = openai_resp_proofs.replace('\n\n', '').strip('\n').split('\n')
    proofs = []
    for paragraph in paragraphs:
        if not paragraph: continue
        without_number = remove_string_identifier(paragraph)
        highlighted_substr = without_number.strip('"').strip("'").strip('.')
        substr_idx_start = -1
        input_text_lower = input_text.lower()
        highlighted_substr_lower = highlighted_substr.lower()
        highlighted_words = highlighted_substr_lower.split(' ')
        
        while len(highlighted_words):
            substr_idx_start = input_text_lower.find(' '.join(highlighted_words))
            if substr_idx_start >= 0:
                break
            highlighted_words = highlighted_words[1:] # remove the leading word of the searched substr (lowercased)
            highlighted_substr = ' '.join(highlighted_substr.split(' ')[1:]) # and from the original substr (not lowercased)
        if substr_idx_start < 0:
             proofs.append(input_text)
        else:
            leading_overlap_start_idx = substr_idx_start - OVERLAP_CHARS if substr_idx_start - OVERLAP_CHARS >= 0 else 0
            leading_overlap = input_text[leading_overlap_start_idx:substr_idx_start]
            trailing_overlap_start_idx = substr_idx_start + len(highlighted_substr)
            trailing_overlap = input_text[trailing_overlap_start_idx:trailing_overlap_start_idx+OVERLAP_CHARS]
            proofs.append(f'{leading_overlap}{SEPARATOR}{highlighted_substr}{SEPARATOR}{trailing_overlap}')
    print('[openai_res_to_references]:\n', proofs)
    return proofs

# Gets the list of objects with questions and answer options,
# returns a list of correct options (to be used for fetching proofs)
def extract_correct_answers(questions):
    res = []
    for question in questions:
        correct_option_idx = question['correct_option_index'] # starts from 0
        correct_option = question['options'][correct_option_idx]
        res.append(correct_option)
    print('[extract_correct_answers]:\n', res)
    return res

# adds the proofs to each question as question[n]['correct_option_index']
# assumes that the number of proofs and their order corresponds to the questions
def add_proofs(proofs, questions):
    for idx, proof in enumerate(proofs):
        questions[idx]['reference'] = proof
    return questions

# Gets the text pasted by the user, does pre-processing (if needed), 
# generates questions using OpenAI API
def get_questions(user_input):
    print('\n\n>>LOADING DATA')
    # temp load a ready json
    # with open('misc/data.json') as f:
    #     data = json.load(f)
    # return data
    openai_questions = openai_get_questions(user_input)
    questions = openai_res_to_questions(openai_questions)
    correct_options = extract_correct_answers(questions)
    openai_proofs = openai_get_proofs(correct_options, user_input)
    proofs = openai_res_to_references(openai_proofs, user_input)
    questions_with_proofs = add_proofs(proofs, questions)
    return { 'questions': questions_with_proofs }

def generate_response_options_states(questions):
    response_options_state = {}
    for idx, question in enumerate(questions, start=1):
        question_index = f'q{idx}'
        response_options_state[question_index] = {}
        for op_idx, _ in enumerate(question['options'], start=1):
            response_options_state[question_index][str(op_idx)] = { 'selected': 0, 'disabled': 0 }
        skip_key = f'skip_{question_index}'
        response_options_state[question_index][skip_key] = { 'selected': 0, 'disabled': 0 }
    return response_options_state

def initialize_state(user_input):
    initilized = 'initilized' in st.session_state and st.session_state['initilized']
    if initilized:
        return
    st.session_state['user_score'] = 0
    questions = get_questions(user_input)
    st.session_state['questions'] = questions
    options = generate_response_options_states(questions['questions'])
    st.session_state['options_state'] = options
    st.session_state['questionnaire_state'] = { 'questions_asked': [], 'questions_skipped': [], 'correct': [] }
    st.session_state['initilized'] = True

def renderQuestions(source='text'):
    def disable_options(question_options_state_keys):
        for question_option in question_options_state_keys:
            question_options_state_keys[question_option]['disabled'] = True
        
    # q_idx - the number part of the key in session_state['options_state'], eg. 1 (to be transformed into 'q1')
    # option_key_to_select - the name of the option key, which needs to be marked as key.selected = true, eg '1' or 'skip_q2'
    def mark_selected(q_idx, option_key_to_select):
        q_index_key = f'q{q_idx+1}'
        for option in st.session_state['options_state'][q_index_key]:
            if option == option_key_to_select:
                st.session_state['options_state'][q_index_key][option]['selected'] = True
            else:
                st.session_state['options_state'][q_index_key][option]['selected'] = False
        
    # q_idx = 0, index in session_state['questions']['questions']
    # q_state_key = {'1': {'selected': 0, 'disabled': 0}, '2': {'selected': 0, 'disabled': 0}, ...}
    # opt_idx_to_select: number = 2
    def mark_selected_disable_all_options(q_idx, q_state_key, option_key_to_select):
        mark_selected(q_idx, option_key_to_select)
        disable_options(q_state_key)

    def highlight_proof(reference):
        # reference is expected to contain only 1 substring, surrounded with separator
        COLOR = '#fea'
        parts = reference.split(SEPARATOR)
        hightlight_tuple = (parts[1], '', COLOR)
        return [
            f'...{parts[0]}', hightlight_tuple, f'{parts[2]}...']
        
    # q_idx - question index (start=1), e.g. 1 for the 1st question (with index 0)
    # answer_status - enum('skipped', 'correct', 'wrong')
    #  st.session_state.questionnaire_state = { 'questions_asked': [], 'questions_skipped': [], 'correct': [] }
    def update_score(q_idx, answer_status='wrong'):
        if q_idx not in st.session_state['questionnaire_state']['questions_asked']:
            st.session_state['questionnaire_state']['questions_asked'].append(q_idx)
            if answer_status == 'skipped':
                st.session_state['questionnaire_state']['questions_skipped'].append(q_idx)
            elif answer_status == 'correct':
                st.session_state['questionnaire_state']['correct'].append(q_idx)
    
    def show_stats():
        questions_asked = len(st.session_state['questionnaire_state']['questions_asked'])
        questions_skipped = len(st.session_state['questionnaire_state']['questions_skipped'])
        answered_correctly = len(st.session_state['questionnaire_state']['correct'])
        divide_by = questions_asked - questions_skipped or 1
        correct_prcnt = answered_correctly / divide_by * 100
        st.markdown(f':grey[Asked: {questions_asked} | skipped: {questions_skipped} | correct: {answered_correctly} ({correct_prcnt}%)]')

    def show_video_proof(proof):
        st.write('Please click "Play" in the video below to listen the relevant segment:')
        start, end = get_proof_time(st.session_state['yt_captions_raw'], proof)
        video_id = get_youtube_video_id(st.session_state['yt_url'])
        yt_link = f'''https://www.youtube.com/embed/{video_id}?start={round(start)}&end={round(end)}&autoplay=1&hl=en&cc_lang_pref=en'''
        print(f'[show_video_proof]: proof = "{proof}", yt_link={yt_link}')
        st.video(yt_link)
    
    def generate_q_options(q_idx, question_with_options, q_idx_started_1, source='text'):
        options = {}
        q_idx_str = str(q_idx_started_1)
        
        # Generate options for UI and 
        # "handlers" to disable all options on click and select only the clicked option
        for idx, option in enumerate(question_with_options['options'], start=1):
            idx_str = str(idx)
            q_key = f'q{q_idx_str}'
            options[idx_str] = st.checkbox(
                option, 
                key=f'{option}_{random.randint(0,1000)}', # we need unique checkboxes
                value = st.session_state.options_state[q_key][idx_str]['selected'], 
                on_change = mark_selected_disable_all_options, args=(q_idx, st.session_state['options_state'][q_key], idx_str,),
                disabled = st.session_state.options_state[q_key][idx_str]['disabled']
                )
        skip_option_key = f'skip_{q_key}'
        options[skip_option_key] = st.checkbox(
            f'Mark question #{q_idx_started_1} as invalid and skip it', 
            value = st.session_state.options_state[q_key][skip_option_key]['selected'], 
            on_change = mark_selected_disable_all_options, args=(q_idx, st.session_state['options_state'][q_key], skip_option_key,),  
            disabled = st.session_state.options_state[q_key][skip_option_key]['disabled']
            ) 
        
        # Logic to check if the option selected is correct, react and update score
        correct_option_idx_key = str(st.session_state['questions']['questions'][q_idx]['correct_option_index']+1)
        
        for key in options:
            value = options[key]
            response_status = 'wrong'
            if value:
                if key == skip_option_key:
                    response_status = 'skipped'
                    st.write(f'☠ Ok, the question #{q_idx+1} was skipped and will be ignored')
                elif key == correct_option_idx_key:
                    response_status = 'correct'
                    st.write('👍 Correct!')
                    annotated_text(*highlight_proof(st.session_state['questions']['questions'][q_idx]['reference']))
                    if source == 'yt':
                        show_video_proof(st.session_state['questions']['questions'][q_idx]['reference'])
                        # st.write('Please click "Play" in the video below to listen the relevant segment:')
                        # start, end = get_proof_time(st.session_state['yt_captions_raw'], st.session_state['questions']['questions'][q_idx]['reference'])
                        # st.video(f'https://www.youtube.com/embed/L8U-pm-vZ4c?start={start}&end={end}&autoplay=1&hl=en&cc_lang_pref=en')
                else:
                    response_status = 'wrong'
                    correct_option_idx = st.session_state['questions']['questions'][q_idx]['correct_option_index']
                    correct_option = st.session_state['questions']['questions'][q_idx]['options'][correct_option_idx]
                    st.markdown(f"😬 Unfortunately, no, the correct answer is *'{correct_option}'*")
                    annotated_text(*highlight_proof(st.session_state['questions']['questions'][q_idx]['reference']))
                    if source == 'yt':
                        show_video_proof(st.session_state['questions']['questions'][q_idx]['reference'])
                        # st.video(st.session_state['youtube_video_url'])
                        # st.write('Please click "Play" in the video below to listen the relevant segment:')
                        # st.video('https://www.youtube.com/embed/L8U-pm-vZ4c?start=60&end=70&autoplay=1&hl=en&cc_lang_pref=en')
                update_score(q_idx+1, response_status)
                show_stats()

    def render_question_title(q_idx):
        st.text("")
        st.subheader(f'''Q{q_idx+1}. {st.session_state['questions']['questions'][q_idx]['question']}''')
    
    for idx, _ in enumerate(st.session_state['questions']['questions']):
        render_question_title(idx)
        generate_q_options(idx, st.session_state['questions']['questions'][idx], idx+1, source)

def show_loader():
    # render the following only if "Generate questions" button was clicked 
    # or questions_rendered flag was set True
    if st.session_state.render_questions_btn or st.session_state.questions_rendered:
        
        # >> Showing the loader (only once after "Generate Q" button click)
        # https://docs.streamlit.io/library/api-reference/layout/st.empty
        processing_done_str = "Processing..... Done"
        with st.empty():
            if st.session_state.questions_rendered:
                st.write(processing_done_str)
            else:
                for seconds in range(2): # this should be dynamic
                    st.write(f"Processing{'.' * seconds}")
                    time.sleep(1)
                st.write(processing_done_str)
        st.session_state.questions_rendered = True

def generate_questions(source='text'):
    input = ''
    if source == 'text':
        input = st.session_state['user_input']
    elif source == 'yt':
        input = st.session_state['user_input_yt']

    initialize_state(input)
    st.header('Questions:')
    renderQuestions(source)

def render_questions_btn_clicked(btn_clicked=False):
    st.session_state['render_questions_btn_clicked'] = btn_clicked
    
def render_questions_btn_yt_clicked(btn_clicked=False):
    st.session_state['render_questions_btn_yt_clicked'] = btn_clicked
    
# Youtube
@st.cache_data
def get_transcript(video_id, lang_code='en'):
  try:
    transcript_list = YouTubeTranscriptApi.list_transcripts(video_id)
    en_transcript = transcript_list.find_transcript([lang_code])
    if not en_transcript:
      print('[getYoutubeCaptions] No English transcript found, halting')
      return None
    transcript = en_transcript.fetch()
    # print(f'''[getYoutubeCaptions] Got Youtube captions for the video {video_id}:
    # {transcript}''')
    print(f'''[getYoutubeCaptions] Got Youtube captions for the video {video_id}''')
    return transcript
  except Exception as e:
    print(f'[getYoutubeCaptions] Error getting youtube captions: ', e)
    return None

def join_yt_transcripts(transcripts):
    text_only = [piece['text'] for piece in transcripts]
    return ' '.join(text_only)

def get_youtube_video_id(youtube_url):
  return youtube_url.split('v=')[1]

def clear_state():
    for key in st.session_state.keys():
        del st.session_state[key]

def get_time_for_proof_start_end(yt_captions, proof, search_start=True):
  # 3 issues with YT proofs: 1) recalculated on each checkbox click, b) error on the last checkbox, c) start/stop are ignored by streamlit, need html container
  highligh = proof.split('<highlight>')[1]
  if not highligh:
    return
  hightlight_processed = highligh.lower()
  highlight_parts = hightlight_processed.split(' ')
  highligh_parts_in_segments_distribution = {}
  for idx, _ in enumerate(highlight_parts):
    searched_phrase = ''
    all_segments_with_hightligh = []
    if search_start:
      if idx == 0:
        searched_phrase = highlight_parts[0]
      else:
        searched_phrase = ' '.join(highlight_parts[0:idx])
    else:
      if idx == 0:
        searched_phrase = highlight_parts[-1:][0]
      else:
        idx_from_end = -1 + idx * -1
        searched_phrase = ' '.join(highlight_parts[idx_from_end:])
    for segment in yt_captions:
      if searched_phrase in segment['text']:
        all_segments_with_hightligh.append(segment)
    highligh_parts_in_segments_distribution[idx] = all_segments_with_hightligh
    # we need a segment which contains the longest highlight part
    # the longer the hightlight part, the less segments will coincide with it
    # and then after adding the next word the hightligh will not be found in any segment
    # so we need the step with 0 found after not-0, and take this not-0 value (expected to be only 1)
    if len(highligh_parts_in_segments_distribution[idx]) == 0 and (idx - 1 >= 0) and len(highligh_parts_in_segments_distribution[idx-1]) > 0:
      the_segment = highligh_parts_in_segments_distribution[idx-1][0]
      if search_start:
        return the_segment['start']
      else:
        segment_ends_at = the_segment['start'] + the_segment['duration']
        return segment_ends_at
  print('[get_time_for_proof_start_end] failed to find start/end for the proof "{proof}"')
  return 0

def get_proof_time(yt_captions, proof):
  start = get_time_for_proof_start_end(yt_captions, proof)
  end = get_time_for_proof_start_end(yt_captions, proof, False)
  last_caption = yt_captions[-1:][0]
  video_length = last_caption['start'] + last_caption['duration']
  YT_SEGMENT_BORDERS_SEC = 3
  start_with_borders = start - YT_SEGMENT_BORDERS_SEC if start - YT_SEGMENT_BORDERS_SEC >= 0 else 0
  end_with_borders = end + YT_SEGMENT_BORDERS_SEC if end + YT_SEGMENT_BORDERS_SEC <= video_length else video_length
  print(f'[get_proof_time] proof="{proof}", start={start_with_borders}, end={end_with_borders}')
  return start_with_borders, end_with_borders
    
### Render UI and combine all together
st.title(PROJECT_NAME)
st.write(PROJECT_IDEA)

tab_text, tab_youtube = st.tabs(['Text', 'Youtube'])

with tab_text:
    user_input = st.text_area(
        TEXT_AREA_INSTRUCTION, 
        DUMMY_TEXT,
        key='user_input',
        max_chars=MAX_INPUT_LEN_CHARS)
    
    # Debug
    # st.write(st.session_state)
    
    st.button(
        'Generate questions', 
        key='render_questions_btn', 
        on_click=render_questions_btn_clicked, args=(True,)
        )
    
    render_questions_btn_was_clicked = 'render_questions_btn_clicked' in st.session_state and st.session_state['render_questions_btn_clicked']
    if render_questions_btn_was_clicked:
        generate_questions()

with tab_youtube:
    url = st.text_input('Youtube video URL')
    
    if url:
        st.session_state['youtube_video_url'] = url
        st.video(url)
        
        st.button(
            'Generate questions', 
            key='render_questions_btn_yt', 
            on_click=render_questions_btn_yt_clicked, args=(True,)
            )
    
        render_questions_btn_yt_was_clicked = 'render_questions_btn_yt_clicked' in st.session_state and st.session_state['render_questions_btn_yt_clicked']
        
        # This block is re-rendered on each checkbox marked
        if render_questions_btn_yt_was_clicked:
            initilized = 'initilized' in st.session_state and st.session_state['initilized']
            if not initilized:
                video_id = get_youtube_video_id(url)
                print(f'\n\nvideo_id={video_id}')
                transcript_obj = get_transcript(video_id)
                transcript_text = join_yt_transcripts(transcript_obj)
                text_first_part = transcript_text[0:MAX_INPUT_LEN_CHARS]
            
                print('setting user_input')
                st.session_state['user_input_yt'] = text_first_part
                st.session_state['yt_captions_raw'] = transcript_obj
                st.session_state['yt_url'] = url
        
            generate_questions('yt')