Voyage AI voyage-3

Voyage AI hakkında

Kuruluş yılı 2023 · Palo Alto, California, USA

Voyage AI was founded in 2023 by Tengyu Ma, an associate professor of Computer Science at Stanford and a recognised researcher on optimisation and representation learning, together with co-founders Dawn Song and Christopher Re among the technical advisory team. The company set out to build best-in-class retrieval and reranking models for RAG, with explicit emphasis on domain-specific variants (code, finance, law, multilingual). Voyage AI raised $20M in seed and Series A funding from CRV, Wing VC, Conviction, AME Cloud, Snowflake and others before being acquired by MongoDB for $220M in February 2025 as the foundational embedding layer for Atlas Vector Search. voyage-3 launched in September 2024 as the company's general-purpose flagship and consistently ranks at or near the top of the MTEB and BEIR retrieval leaderboards.

Mimari

Transformer bi-encoder with task-specific instruction-tuning and Matryoshka heads

Voyage AI voyage-3 is a hosted general-purpose embedding model with a 32,000-token context window, the longest commercial embedding context as of late 2024. The default output is a 1,024-dim vector, but the API also supports 256-, 512- and 2,048-dim heads via Matryoshka-style training, all sharing the same backbone. Training used a contrastive retrieval objective over a large curated multilingual mix including code, scientific and financial documents; the team has not disclosed exact token counts. Voyage exposes an input_type parameter ('query' or 'document') for asymmetric search and a separate output_dtype parameter for int8 / binary quantisation, reducing storage cost by up to 32x with a small quality drop. voyage-3 is bundled with a paired reranker (rerank-2) for two-stage retrieval. The model is offered exclusively via hosted API and is also natively integrated into MongoDB Atlas Vector Search since the February 2025 acquisition.

Parametreler

Undisclosed

Bağlam

32.000 token

Yetenekler

• 32,000-token context window (longest commercial embedding context)
• Matryoshka-style heads: 256 / 512 / 1024 / 2048 dimensions
• input_type parameter for asymmetric query / document retrieval
• int8 and binary quantisation for cheap vector storage
• Multilingual coverage with strong English and code performance
• Top-tier MTEB and BEIR retrieval scores
• Native MongoDB Atlas Vector Search integration
• Best for: production RAG, long-document retrieval, MongoDB-backed apps

Eğitim ve lisans

Not disclosed. Voyage describes a 'large curated multilingual mix including code, scientific and financial documents' plus contrastive negatives.

Lisans: Proprietary commercial API. Available standalone and bundled with MongoDB Atlas Vector Search.

Güvenlik testleri: Embeddings are not subject to content-filter constraints. Voyage publishes general fairness statements.