• Stars
    star
    188
  • Rank 205,563 (Top 5 %)
  • Language
    Python
  • License
    MIT License
  • Created about 1 year ago
  • Updated 4 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Semantic search for competitive programming problems

Is my problem new?

A simple semantic search engine on competitive programming problems.

Demo server (while supplies last...) [Buy me a boba]

The included problems of Codeforces & BZOJ are for fair uses only. Their copyrights belong to the original authors and the LICENSE included only apply for the source codes included in this repository.

Screenshots

How does this work?

This idea is simple:

  1. Simplify the statement & remove background by prompting chatgpt.

  2. Embed the simplified documents and queries to perform vector searches.

It only happens recently that both models are good and cheap enough.

This pipeline is also not limited, of course, to competitive programming problems. You can use it to search for any kind of documents by modifying the prompt.

Deploy

Disclaimer: This project was finished in ~2 days so do not expect high quality code. Deploy at your own risk. If you want to serve it publicly, add safety measures such as rate limits, authentication, etc.

You will need API keys from OpenAI (https://platform.openai.com/) and Cohere (https://cohere.com/). The model gpt-3.5-turbo (also known as chatgpt) from OpenAI and cohere-embed-english-v3.0 from Cohere (empirically works better than OpenAI's embeddings at the same price) are used. You can check their pricings online.

  1. Copy settings_sample.json to settings.json. Fill in the API keys.

  2. Download embeddings from here and place it in the folder (so that embs/embs_cohere.npy exists) or run python -m src.build_embedding.py (regenerate embeddings using cohere, costs ~$0.5).

  3. Run python -m src.ui.py to start your server!

Can we make this completely open-source? You will need to replace the summary model and the embedding model with open source versions. For embedding models, this leaderboard may be a good place to look for.

Adding a new set of problems

  1. Generate a json file in problems/ folder. You should supply uid, url, tag and statement fields. See problems/format.md for detailed specifications and see src/scrapper/ for two examples: a crawler for Codeforces and a parser for bzoj (using a local dump).

  2. Generate summaries with ChatGPT. Run python -m src.build_summary.py.

  3. Generate embeddings with Cohere. Run python -m src.build_embedding.py.

For reference, adding all problems from codeforces & bzoj costs about $20.

Known issue: The OpenAI api has a rate limit and to make sure our requests fit in the limit, currently the code sleeps after requests following to the rate limit. You may want to adjust the logic to increase throughput.