Building a RAG Pipeline for Semantic Code Search

As you would hope (or expect) from JetBrains people, there's actually quite a lot of interesting detail about how exactly they use embeddings for code search. Namely:

• Sometimes quantization but never (?) Matryoshka
• When you just need to be in the right neighborhood of correct code then you can drop to binary precision but if you need a yes/no "related code" decision then you often need to preserve more precision
• They use their fancy AST tools to determine chunks