Skip to content

Latest commit

 

History

History
72 lines (52 loc) · 2.71 KB

README.md

File metadata and controls

72 lines (52 loc) · 2.71 KB
Raw

Infinity Quiz

inf-quiz

Description

Use OpenAIs GPT-3.5 to generate an infinite trivia questions and answers.

Engineer prompts from a list of trivia categories, questions and answers (as "few-shots"). Let the LLM reply with json format. Use the generated prompts to generate more trivia questions and answers. Repeat.

How to use?

Install the requirements:

pip install -r requirements.txt

Create a .env file in src with your OpenAI API key:

# Development settings
OPENAI_API_KEY=YOUR_API_KEY
SECRET_KEY=ANY_SECRET_KEY_STRING

Run the server:

python app.py

Then go to localhost (default port is 5000) and start generating trivia questions and answers.

Prompt Engineering

The prompt is designed as following:

Generate a new question, in the following (category, subcategory):  

# Randomised from a list of categories and subcategories.
(Technology, Computers and Internet)

# Randomised from a list of example questions and answers.
[{"Question": "Which city is the capital of Australia?", "Options": ["Sydney", "Melbourne", "Canberra", "Perth"], "Answer": "Canberra"}, {"Question": "What planet is known as the 'Red Planet'?", "Options": ["Venus", "Mars", "Jupiter", "Saturn"], "Answer": "Mars"}, {"Question": "Who wrote the novel 'To Kill a Mockingbird'?", "Options": ["Harper Lee", "J.D. Salinger", "George Orwell", "Ernest Hemingway"], "Answer": "Harper Lee"}] 

Provide a json response below:

The response from the example above was:

{
    'Question': "What does the acronym 'HTML' stand for?", 
    'Options': ['Hyper Text Markup Language', 'Hyperlinks and Text Markup Language', 'Hipster Text Messaging Language', 'High Tech Media Language'], 
    'Answer': 'Hyper Text Markup Language'
}

I do not enforce the json schema, and completely rely on the LLMs ability to understand the requirements.

I also ask for a rationale, depending on the user's selected option. The rationale has a "mood" parameter, generated from data/moods.json.

TODO

  • include category and subcategory in the prompt
  • Keep a stack of three questions all the time to avoid latency.
  • Add catchy rationale (with source?) to each answer (and option?).
    • Rationale could be generated from the correct question + selected option.
  • Store all selected options in a database.
  • Rate questions and answers.
  • Wrap the generator in a object, so i can call it by QuizGenerator(category, subcategory, difficulty).
  • Do it about vett och etikett.
  • include a hint in the response.
  • add a database to visualise how many questions have been generated, which questions are the most appreciated, and how many incorrect json responses there have been