GitHub Bolsters Copilot with Smarter Embedding Model for VS Code

GitHub has rolled out a major upgrade to its Copilot tool, introducing a new embedding model that makes code search within VS Code faster, more accurate, and more memory-efficient.

The update is meant to sharpen how Copilot retrieves the most relevant code snippets and documentation in response to developer queries.

Under the hood, embeddings are vector representations that help line up developer queries with semantically related code and natural language, even if exact keywords differ.

"After listening to the community’s feedback, today we are rolling out a new GitHub Copilot embedding model that makes code search in Visual Studio Code faster, lighter on memory, and far more accurate. This means retrieving the snippets you actually need instead of near misses.

"It delivers a 37.6% lift in retrieval quality, about 2x higher throughput, and an 8x smaller index size, so GitHub Copilot chat and agentic responses are more accurate, results return faster, and memory use in VS Code is lower," Asha Sharma, President, CoreAI Product at Microsoft said.

The new model GitHub deployed is trained especially for code and documentation contexts, powering retrieval in various Copilot modes including chat, edit, and agent workflows.

GitHub reports that this upgrade yields a 37.6 % lift in retrieval quality, approximately double the throughput, and an eight-fold reduction in index size. In real usage, this means faster responses, lower memory usage in VS Code, and more precise matches instead of near misses. For example, Copilot is now better at distinguishing between semantically close code variants (e.g. findOne vs find) and recommending the correct one.

GitHub details that, in tests, C# developers saw a 110.7% increase in code acceptance rates and Java developers saw 113.1%, reflecting stronger confidence in the suggestions. The retraining strategy involved “hard negatives” (incorrect yet plausible alternatives) to teach the model finer distinctions.

Looking ahead, GitHub plans to broaden language support, refine negative mining, and leverage efficiency gains to scale even larger, more precise models. The aim: make AI coding assistance increasingly reliable and context-aware for developers everywhere.