Voyage AI

Voyage AI is a company specializing in state-of-the-art embedding models and rerankers for semantic search, retrieval-augmented generation (RAG), and agentic applications. Their models are designed for high accuracy, low dimensionality, low latency, and long-context capabilities.

To get started with Voyage AI, you need an API key from Voyage AI. You should sign in to your Voyage AI account and create a new API key, which you can then use to authenticate requests to the Voyage AI API.

The main features of Voyage AI include Embedding Models, which generate vector embeddings for text optimized for retrieval, search, and semantic similarity; Rerankers, which rank documents by relevance to a query for improved search results; and various Models, with multiple options available for different use cases such as general, finance, law, code, and multilingual applications.

The available embedding models include voyage-3-large, which offers 1,024 (default), 256, 512, or 2,048 dimensions and a 32,000 max token limit, and is the best general-purpose and multilingual retrieval quality model. Voyage-3 has 1,024 dimensions and a 32,000 max token limit, optimized for general-purpose and multilingual retrieval. Voyage-3-lite features 512 dimensions and a 32,000 max token limit, optimized for latency and cost. For code retrieval, voyage-code-3 is available with 1,024 (default), 256, 512, or 2,048 dimensions and a 32,000 max token limit. Voyage-finance-2, with 1,024 dimensions and a 32,000 max token limit, is optimized for finance retrieval and RAG. Voyage-law-2, featuring 1,024 dimensions and a 16,000 max token limit, is optimized for legal retrieval and RAG. Lastly, voyage-code-2, an previous generation model, offers 1,536 dimensions and a 16,000 max token limit and is optimized for code retrieval.

To use the embedding models, you need to specify the model name and API key. You can optionally set the input type to query, document, or None, and the API will then return embeddings for your text.

The input type parameter can be set to query, which is used for search or retrieval queries and causes the model to prepend a prompt to optimize for query use cases. Alternatively, document is used for documents or content you want to be retrievable, with the model prepending a prompt to optimize for document use cases. If set to None, the input text is directly encoded without any additional prompt.

Voyage AI has rate limits on API usage. Check the official documentation for specific limits and tips on how to avoid hitting them.

The API supports truncation. If your text exceeds the model's maximum token limit, you can set the truncation parameter to true to automatically truncate the input.

You can find more information in the official Voyage AI documentation at https://docs.voyageai.com, through community providers and integrations at https://ai-sdk.dev/providers/community-providers/voyage-ai, or by checking the LangChain integration details at https://docs.langchain.com/oss/javascript/integrations/text_embedding/voyageai.