voyage-code-3

Voyage-code-3 is a domain-specialized embedding model that belongs to Voyage's family of targeted retrieval solutions, alongside models built for finance, legal, and general-purpose tasks. Rather than serving as a general-purpose text embedder, it is engineered specifically to represent code artifacts in a high-dimensional space where semantically similar code segments cluster together, enabling downstream systems to pull the right context during tasks like code completion, search, and automated review. The model supports multiple embedding dimensions, allowing developers to trade off precision against storage and compute costs, and includes support for embedding quantization—an important practical feature when deploying code retrieval pipelines at scale.

The lineage of voyage-code-3 reflects an iterative refinement process within Voyage's domain-specific embedding program, with the prior voyage-code-2 already demonstrating a 17% improvement over alternative solutions at its release. This progression suggests a deliberate effort to study code structure, naming conventions, and cross-file dependencies during training, building embeddings that are robust to variable naming and functional equivalence. Developers integrating this model typically feed code snippets, documentation, or function signatures into retrieval pipelines, relying on the model's ability to surface relevant blocks even when surface-level wording differs. The model runs in a managed API environment with configurable rate limits that scale with usage tier, making it practical for both prototyping and production-grade code intelligence features.