Voyage AI

Voyage AI sits inside the retrieval layer of a RAG pipeline — between your raw documents and your LLM. The workflow the vendor describes is conventional: chunk your unstructured data, generate embeddings via the API, store them in a vector database of your choosing, and then optionally run a reranker against the top candidates before passing context to the language model. The service is API-only and proprietary, so integration is a straightforward HTTP call from any stack, but the compute and data handling happen entirely on Voyage AI’s infrastructure.

The clearest differentiator is the model tier structure. General-purpose models handle multilingual retrieval out of the box. Domain-specific models — the vendor names finance, legal, and code as explicit targets — are optimized for vocabularies and query patterns that trip up general embeddings. Company-specific fine-tuning is offered as a separate track for teams whose internal terminology diverges enough from public corpora that off-the-shelf models return irrelevant results. The reranker lineup, including the rerank-2.5 and rerank-2.5-lite variants, adds instruction-following capability, meaning you can bias ranking behavior without retraining the embedding model.

The 32K-token context window — which the vendor describes as the longest available in commercial embedding — matters when your source documents are long and chunking aggressively enough to fit a smaller window loses critical cross-chunk relationships. The voyage-context-3 model is specifically described as capturing chunk-level details while retaining global document context, addressing a common failure where chunk-isolated embeddings miss meaning that spans sections. Where this setup breaks: teams with strict data residency requirements, air-gapped environments, or budgets that require self-hosted inference cannot use this service at all — there are no open weights and no on-premises deployment path.