PulseGym members ask the same questions all day: prices, freezing, guests, class rules. Your job: build an assistant that answers them correctly, only from the gym's rules, and knows when to say "I don't know".
What you build
An assistant that answers questions using only gym-rules.md, served through the API in API-CONTRACT.md.
It must:
- Answer only from the rules, and cite the rule IDs it used (for example R-07).
- Use the newer rule when an Update changes an old one (the Updates section overrides older rules).
- Combine two rules when a question needs both.
- Say clearly that it doesn't know when the rules don't cover the question, instead of making something up.
- Answer in the language of the question: Arabic or English.
- Refuse medical or injury advice and refer the member to a doctor and a trainer.
How we run it
- Your repo has a
Dockerfileand adocker-compose.yml. We rundocker compose up --buildand callhttp://localhost:8080. - Before you submit, run
python3 check.py. It sends the 10 sample questions insample-questions.jsonand shows your answers next to the expected ones.
Our hidden test sends 30 different questions: simple ones, ones that need two rules, ones affected by Updates, Arabic questions, questions the rules don't cover, and medical questions.
Models
Any model or library is fine. Free options are enough: the Gemini API free tier, the Groq free tier, or a local model with Ollama. Paid models earn no extra points. Read any API key from an environment variable and say so in your README.
What you submit
- A GitHub link (public, or shared with us) or a zip of your code.
- A
README.md(one page): your approach (how you find the right rules, how you prevent made-up answers), its limits, and how you used AI tools. - Your answers to the 10 sample questions.
How you are scored (out of 100, same sheet for all tracks)
| Criteria | Points |
|---|---|
| Passes the hidden test | 35 |
| Quality of the approach | 25 |
| Edge cases and thinking ahead | 15 |
| Clear explanation (README) | 15 |
| Presentation (easy to run and read) | 10 |
Answers are judged on meaning, not exact wording. A made-up rule or a medical tip counts against you more than a wrong answer.
Tips
- A good answer that cites the wrong rule is still a weak answer. Make the citations real.
- "I don't know" is a correct answer when the rules are silent.
- Test in Arabic, not only English.