How codesearch.ai works

Welcome to the third issue of Tour de Source, a newsletter that dives into the source code of great open source projects and how they work!

This week our guest Rok Novosel will be giving a deep dive into the source of codesearch.ai, an experimental AI-powered code search engine that answers natural language queries with functions indexed from GitHub.com and StackOverflow.

It uses Hugging Face Transformers under the hood, and the training procedure is inspired by a paper called Text and Code Embeddings by Contrastive Pre-Training from OpenAI. The CodeSearchNet project served as a basis for data collection and cleaning.

See how codesearch.ai works!

Next Steps

• 💬 Discuss this issue in our Discord (#tour-de-source)

• 🤔 Let us know what you thought on Twitter

• 🔍 Browse the code

• 👩‍💻 See Sourcegraph in action! Schedule time with an engineer.