Skip to content
All Models

Voyage Code 2

voyage/voyage-code-2

Voyage Code 2 is Voyage AI's code-specialized embedding model with a context window of 0 tokens. It improves code retrieval by 14.52% over OpenAI text-embedding-3-large and supports Python, C++, Java, and major ML framework documentation.

index.ts
import { embed } from 'ai';


const result = await embed({
  model: 'voyage/voyage-code-2',
  value: 'Sunny day at the beach',
})

What To Consider When Choosing a Provider

  • Zero Data Retention

    AI Gateway does not currently support Zero Data Retention for this model. See the documentation for models that support ZDR.

    Authentication

    AI Gateway authenticates requests using an API key or OIDC token. You do not need to manage provider credentials directly.

Voyage Code 2 targets code search. If you're embedding source code, function signatures, and documentation for retrieval, it outperforms general-purpose embedding models by a wide margin.

Voyage AI released voyage-code-3, which supports 300+ programming languages, a 32K context window, and Matryoshka dimensionality. Use voyage-code-3 for new deployments unless you need compatibility with existing Voyage Code 2 indices.

Despite its code focus, Voyage Code 2 outperforms several general-purpose models on standard text retrieval. Use it for mixed code-and-documentation corpora without a second model.

When to Use Voyage Code 2

Best For

  • Code search engines:

    Retrieve relevant functions, classes, or modules from natural language queries

  • Code completion pipelines:

    Retrieval-augmented generation finds similar code patterns

  • Developer documentation search:

    API references, library docs, and code examples

  • Mixed code and text retrieval:

    A single model handles both source code and natural language documentation

  • ML framework documentation:

    Retrieval for Python-centric data science and machine learning workflows

Consider Alternatives When

  • You need broader language coverage:

    Voyage-code-3 supports 300+ programming languages beyond Python, C++, and Java

  • You need a longer context window:

    Voyage-code-3 offers 32K tokens versus Voyage Code 2's 0 tokens

  • Your workload is general-purpose text with no code:

    A general-purpose embedding model like voyage-3.5 fits better

  • You need Matryoshka dimensionality:

    Voyage-code-3 supports 2048/1024/512/256 dimensions for flexible sizing

Conclusion

Voyage Code 2 delivers a 14.52% code retrieval improvement over OpenAI text-embedding-3-large. If you have existing Voyage Code 2 indices, you can keep them and avoid a re-embed. For new deployments, use voyage-code-3 for its broader language coverage, longer context window, and Matryoshka dimensionality. Route requests through AI Gateway for unified access.

FAQ

Voyage Code 2 is evaluated on Python, C++, and Java, along with documentation for major ML libraries (NumPy, Pandas, PyTorch, scikit-learn, TensorFlow, Matplotlib, SciPy). It handles both text-to-code and code-to-code retrieval.

Voyage Code 2 achieves a 14.52% improvement in recall@5 over OpenAI text-embedding-3-large across 11 code retrieval datasets derived from HumanEval, APPS, MBPP, DS-1000, CodeChef, and LeetCode.

For new deployments, voyage-code-3 is the recommended choice. It supports 300+ programming languages (vs. Python, C++, Java), a 32K context window (vs. 16K), and flexible Matryoshka dimensions. Use Voyage Code 2 if you have existing indices and want to avoid re-embedding.

Yes. Voyage Code 2 exceeds OpenAI text-embedding-3-large by 3.03% on general-purpose text retrieval, making it viable for mixed code-and-documentation corpora.

0 tokens. This handles most individual source files and documentation pages. For larger code contexts, use voyage-code-3 with its 32K-token window.

Add your Voyage AI API key in AI Gateway settings, then send embedding requests through AI Gateway. AI Gateway authenticates requests across providers.

11 code retrieval datasets totaling 43,909 query-document pairs, derived from HumanEval, APPS, MBPP, DS-1000, CodeChef, and LeetCode. The evaluation metric is recall@5.