Why it matters
- Consistently top MTEB benchmark performance — objectively the best embedding quality available via API for many retrieval tasks.
- Domain-specific models (code, finance, legal) are not available from OpenAI or Cohere — a unique capability.
- Reranking API enables two-stage retrieval that significantly improves RAG precision over single-stage embedding search.
- Used by Anthropic — Claude's documentation Q&A and RAG features use Voyage AI embeddings.
Key capabilities
- voyage-3 (general): State-of-the-art general embedding model; top MTEB performance.
- voyage-code-2: Domain-specific embedding for code search, code Q&A, and technical retrieval.
- voyage-finance-2: Optimized for financial document retrieval (10-Ks, earnings calls, market research).
- voyage-law-2: Legal document retrieval (contracts, case law, regulations).
- voyage-multilingual-2: 100+ language support with strong cross-lingual retrieval.
- Rerankers: rerank-1 and rerank-lite-1 for post-retrieval precision improvement.
- Instruction following: Embedding models that accept task instructions for better retrieval vs. classification.
- REST API: Simple POST endpoint; Python and TypeScript SDKs.
Technical notes
- Models: voyage-3, voyage-3-lite, voyage-code-2, voyage-finance-2, voyage-law-2, voyage-multilingual-2
- Dimensions: 1024 (voyage-3); 512 (voyage-3-lite) — configurable
- Max tokens: 32,000 tokens per input (voyage-3)
- API: REST; Python SDK (
pip install voyageai); TypeScript SDK - Pricing: Pay-per-token; ~$0.06/1M tokens (voyage-3); free tier available
- Company: Voyage AI; Stanford-affiliated; founded 2023 by Tengyu Ma and colleagues
Ideal for
- RAG applications in specialized domains (legal, finance, code) where domain-specific embeddings significantly improve retrieval.
- Teams who have measured their RAG performance and need the highest possible retrieval accuracy.
- Applications needing two-stage retrieval with reranking for precision-critical use cases.
Not ideal for
- Simple general-purpose semantic search where OpenAI or Cohere embeddings are sufficient.
- Teams heavily invested in OpenAI's ecosystem who want tighter integration.
- Large-scale embeddings on a tight budget — premium quality comes at premium pricing.
See also
- Cohere Embed — Strong competitor with better enterprise cloud marketplace availability.
- Pinecone — Vector database for storing and searching Voyage AI embeddings.
- Jina Embeddings — Open-source-first embedding models with 8K token context window.