voyage-code-2 generates embeddings for code by converting source code and code-related text into fixed-length numeric vectors that capture semantic intent rather than surface syntax. From a developer’s perspective, this means you send a block of code (for example, a function, class, or configuration snippet) to the model, and it returns a vector representation. Code that performs similar tasks—such as two different implementations of request retries or authentication checks—will tend to produce vectors that are close together, even if variable names, formatting, or language constructs differ. You do not need to parse the code into an abstract syntax tree or annotate it manually; the model handles the representation internally.
The model works best when code is embedded in meaningful, coherent units. In practice, that usually means embedding one function, class, or small module at a time, rather than entire repositories or very large files. Developers often prepend contextual information—like file path, function signature, or short comments—to the code before embedding. This extra context helps the embedding reflect what the code is for, not just how it is written. voyage-code-2 treats code and code-adjacent text (comments, docstrings) as part of the same semantic input, which improves retrieval quality when users search with natural language.
Once embeddings are generated, they are typically stored in a vector database such as Milvus or Zilliz Cloud. The model itself does not manage storage or search; it only produces vectors. The vector database indexes those vectors and enables similarity search at scale. This separation is important: voyage-code-2 focuses on representation quality, while Milvus or Zilliz Cloud handles indexing, filtering, and fast retrieval across large code collections.
For more information, click here: https://zilliz.com/ai-models/voyage-code-2