Voyage AI

What Voyage AI does

Voyage AI provides cutting-edge embedding models and rerankers to supercharge retrieval for unstructured data. Embedding models convert unstructured data (documents, images, audios, videos, tabular data) into dense numerical vectors (embeddings) that capture semantic meanings, serving as building blocks for semantic search and Retrieval-Augmented Generation (RAG). Rerankers are neural nets that output relevance scores between a query and multiple documents, refining initial retrieval results for more accurate relevancy. Voyage AI offers API endpoints for these models, which seamlessly integrate with other RAG stack components like vector stores and Large Language Models (LLMs).

Products

• Voyage 4 Model SeriesBest-in-class embedding models and rerankers, including voyage-multimodal-3.5.
• voyage-4-largeBest general-purpose and multilingual retrieval quality embedding model with a context length of 32,000 tokens and embedding dimensions of 1024 (default), 256, 512, 2048. All embeddings created with the 4 series are compatible with each other.
• voyage-4Optimized for general-purpose and multilingual retrieval quality embedding model. All embeddings created with the 4 series are compatible with each other.
• voyage-4-liteOptimized for latency and cost embedding model. All embeddings created with the 4 series are compatible with each other.
• voyage-code-3Optimized for code retrieval embedding model.
• voyage-finance-2Optimized for finance retrieval and RAG embedding model with a context length of 1024 tokens.
• voyage-law-2Optimized for law retrieval and RAG embedding model with a context length of 16,000 tokens, also improved performance across all domains.
• voyage-code-2Optimized for code retrieval (17% better than alternatives) / Previous generation of code embeddings with an embedding dimension of 1536.
• voyage-3-largePrevious generation of text embeddings for general-purpose and multilingual retrieval quality.
• voyage-3.5Previous generation of text embeddings optimized for general-purpose and multilingual retrieval quality.
• voyage-3.5-litePrevious generation of text embeddings optimized for latency and cost.
• voyage-3Optimized for general-purpose and multilingual retrieval quality embedding model.
• voyage-3-liteOptimized for latency and cost embedding model with an embedding dimension of 512.
• voyage-multilingual-2Multilingual embedding model.
• voyage-large-2-instructInstruction-tuned general-purpose embedding model optimized for clustering, classification, and retrieval. Top of MTEB leaderboard.
• voyage-large-2General-purpose embedding model optimized for retrieval quality (e.g., better than OpenAI V3 Large).
• voyage-2General-purpose embedding model optimized for a balance between cost, latency, and retrieval quality with a context length of 4000 tokens.
• voyage-lite-02-instructInstruction-tuned for classification, clustering, and sentence textual similarity tasks.
• voyage-02Pilot-version v2 embedding model.
• voyage-01V1 embedding model.
• voyage-lite-01Lite v1 embedding model.
• voyage-lite-01-instructTweaked on top of voyage-lite-01 for classification and clustering tasks.
• voyage-4-nanoOpen-weight embedding model available on Hugging Face. All embeddings created with the 4 series are compatible with each other.
• rerank-2.5Generalist reranker optimized for quality with instruction-following and multilingual support, with a context length of 32,000 tokens.
• rerank-2.5-liteGeneralist reranker optimized for both latency and quality with instruction-following and multilingual support.
• rerank-2Generalist second-generation reranker optimized for quality with multilingual support, with a context length of 16,000 tokens.
• rerank-2-liteGeneralist second-generation reranker optimized for both latency and quality with multilingual support, with a context length of 8000 tokens.
• rerank-1Generalist first-generation reranker optimized for quality with multilingual support.
• rerank-lite-1Generalist first-generation reranker optimized for both latency and quality, with a context length of 4000 tokens.
• voyageai Python packagePython library to access Voyage AI text embedding and reranker models via API.

Key capabilities

• Best-in-class embedding models and rerankers
• High accuracy in retrieving relevant contextual information
• Low dimensionality (3x-8x shorter vectors)
• Low latency (4x smaller model and faster inference)
• Cost efficient (2x cheaper inference)
• Long-context (longest commercial context length available - 32K tokens)
• Modularity (plug-and-play with any vectorDB and LLM)
• Deploy Anywhere
• General-purpose models
• Domain-specific models
• Company-specific models
• Instruction-following and multilingual support for rerankers
• API endpoints for embedding and reranking models
• Python API for easy integration

Use cases

• RAG retrieval and response quality
• Semantic search
• Retrieval-Augmented Generation (RAG)
• Domain-specific chatbots
• Company-specific chatbots
• AI applications
• Clustering
• Classification
• Sentence textual similarity tasks
• Code retrieval
• Finance retrieval
• Law retrieval

AI approach

Voyage AI provides cutting-edge embedding models and rerankers. Embedding models convert unstructured data into dense numerical vectors (embeddings) that capture semantic meanings, essential for semantic search and Retrieval-Augmented Generation (RAG). Rerankers output relevance scores between a query and multiple documents to refine retrieval results. Voyage AI offers API endpoints for these models.

Tech named: embedding models, rerankers, neural net models, transformers, vectorDB, LLM, cross-encoders, BM25, TF-IDF, tokenizer

Industries served

• Finance
• Legal

What it says sets it apart

• State-of-the-art in retrieval accuracy
• Longest commercial context length available (32K tokens)
• Optimized for specific domains like finance, legal, and code
• Offers open-weight models (voyage-4-nano)
• Provides both embedding models and rerankers for a complete retrieval solution
• High accuracy, low dimensionality, low latency, and cost efficiency

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.