Developers use voyage-code-2 for semantic search by embedding both code artifacts and search queries into the same vector space, then retrieving the closest matches by similarity. The typical workflow starts with ingestion: code is broken into retrievable units (functions, classes, config blocks), each unit is embedded using voyage-code-2, and the resulting vectors are stored alongside metadata. At query time, a user’s natural-language question or code snippet is embedded and compared against stored vectors to find the most relevant matches.
This approach enables searches that keyword systems struggle with. For example, a developer might search for “where do we throttle API requests,” even though the code uses terms like rate_limiter or quota_guard. Semantic search powered by voyage-code-2 can retrieve the correct implementation because the embeddings reflect intent rather than exact wording. Developers often combine similarity search with metadata filters, such as restricting results to a specific repository, service, or language.
The search experience becomes production-ready when embeddings are stored in a vector database such as Milvus or Zilliz Cloud. These systems handle indexing and fast nearest-neighbor search, even when the codebase contains millions of embedded units. The result is a semantic search system that feels responsive and reliable, rather than a slow or approximate prototype.
For more information, click here: https://zilliz.com/ai-models/voyage-code-2