voyage-code-2 works well across most mainstream programming languages, especially those commonly found in real-world production systems. Languages such as Python, JavaScript, TypeScript, Java, Go, and C++ tend to work particularly well because they are widely represented in modern codebases and are commonly paired with rich comments and documentation. That said, voyage-code-2 does not rely on language-specific rules; it focuses on patterns of structure, naming, and behavior that generalize across languages.
In practice, the “best” language is less important than how the code is chunked and contextualized. A clearly scoped function with a descriptive name and short comment will usually embed better than a large file containing many unrelated responsibilities, regardless of language. For example, embedding a single validate_token() function in Python or a TokenValidator class in Java tends to produce clean, searchable vectors. Embedding an entire 2,000-line file with multiple concerns usually produces noisy embeddings that are harder to retrieve accurately.
For multilingual repositories, voyage-code-2 can be especially useful when combined with metadata and vector search. By storing embeddings in a vector database such as Milvus or Zilliz Cloud, developers can filter by language, repository, or module before running similarity search. This ensures that a query like “retry HTTP request” returns relevant results in the expected language or service. The model provides cross-language semantic understanding, while the vector database enforces practical constraints.
For more information, click here: https://zilliz.com/ai-models/voyage-code-2