GitHub Copilot gets smarter at finding your code: Inside our new embedding model

GitHub Copilot's new embedding model significantly improves code search in VS Code, delivering 37.6% better retrieval quality, 2x higher throughput, and 8x smaller index size. This helps developers find exactly the code snippets they need, with C# and Java developers seeing over 110% improvement in code acceptance. The model was trained using contrastive learning with "hard negatives" - code examples that look correct but aren't - to distinguish between relevant and nearly relevant code. These improvements make GitHub Copilot chat more accurate while reducing memory usage and speeding up results.