Building a RAG Pipeline for Semantic Code Search: A Developer Diary and Field Notes

Somewhere deep inside a massive code repository, an AI agent is hunting for the function that refreshes session tokens. It won't find the word "refresh" anywhere in the code. That's exactly the problem JetBrains set out to solve.

The company recently detailed how it built JetBrains Context, a semantic code search platform that helps AI coding agents find relevant code by meaning, not just keywords. The system uses a technique called retrieval-augmented generation, or RAG, breaking source files into carefully scoped chunks using language-aware parsers, then converting those chunks into mathematical vectors that cluster similar concepts together.

One of the trickier engineering choices involved storage. Millions of code chunks produce millions of vectors, each potentially weighing 16 kilobytes. JetBrains found that compressing each vector down to a single bit per dimension, 32 times smaller than full precision, preserved retrieval quality surprisingly well.

Crucially, the system never stores actual source code on JetBrains servers, and embeddings are generated on JetBrains-operated hardware, so no proprietary code touches outside vendors.