Memorate is an application to generate multiple choice questions to a given text or Youtube video, indented to help with the memorization of information. This is a pet project, which I created getting acquainted with the OpenAI API (and also streamlit.io).
# clone the repo
git clone [email protected]:IuriiD/memorate.git
cd memorate
# create the virtual environment and install the dependencies (commands may differ depending on your setup)
python3 -m venv env
source env/bin/activate
pip3 install -r requirements.txt
# create keys.py (see keys_example.py) and provide your OPENAI_KEY, OPENAI_ORG
# run the app, will start at http://localhost:8501/
streamlit run index.py
The UI was created using Streamlit, a popular open-source Python library for building data science web applications. The app has 2 tabs: "Text" and "Youtube", where some text or a Youtube video URL can be pasted, correspondingly. After clicking "Generate Questions", user is presented with 3 multiple choice questions. Each question has 3 options + an option to mark the question as irrelevant.
When some option is chosen (besides the one that marks the question as irrelevant), the user receives feedback, including the correct answer and a reference to the source (phrase/paragraph in the text or the part of the YouTube video where this question is answered). Also, the app calculates the % of correct answers (excluding the questions marked by the user as irrelevant).
The app was written on Python (mainly to use Streamlit for quick UI). I'm a node.js developer, so I apologize for the possible straightforwardness of some implementations. OpenAI completions endpoint was used, model text-davinci-003, temperature=0.1, max_tokens=400, otherwise default parameters. The logic was implemented using vanilla Python, without any frameworks like Langchain etc. Most of the code is about parsing the model's outputs and rendering the questionnaire in Streamlit (which as it appeared is not really intended for making questionnaires).
The app makes 2 requests to the completions endpoint:
- one to generate questions with response options and
- another to find the relevant places in the original text which confirm the correct answers.
The app is a proof-of-concept, covers only the "happy flows" and is not claimed to be production ready in any way ;)
- Other types of questions could be generated (yes/no, single-choice, fill-in-blanks, open-ended, choose the best summary, check if summary is correct etc)
- After answering the questions, if the user makes mistakes, they can answer the problematic questions, rephrased, till no mistakes are made
- Optionally, the user may get these (all or problematic) questions asked to them again (sent via email or Facebook messenger bot etc - see SuperMemo2 algorithm)