Voyage AI

Converts unstructured text into dense vector representations using the voyage-3.5 model, supporting up to 32K tokens per input—the longest commercial context window available. The model is optimized for semantic similarity and retrieval tasks, producing 3x-8x shorter vectors than competing embeddings while maintaining or exceeding accuracy on standard benchmarks. Vectors can be directly indexed into any vector database without preprocessing or dimensionality reduction.

Provides voyage-3.5-lite, a smaller variant optimized for inference speed and memory efficiency without significant accuracy degradation. Designed for edge deployment, mobile applications, or high-throughput batch processing where latency and computational cost are primary constraints. Maintains compatibility with standard vector database APIs while reducing per-request inference time.

Voyage embeddings produce 3x-8x shorter vectors compared to competing embeddings (OpenAI, Cohere) through undisclosed dimensionality reduction or quantization techniques. Shorter vectors reduce vector database storage costs, index size, and search latency without sacrificing retrieval accuracy. Enables cost-effective scaling of large-scale RAG systems and semantic search applications.

Provides specialized embedding models fine-tuned on domain-specific corpora (finance documents, legal contracts, source code) to improve semantic understanding and retrieval accuracy within those domains. Models are trained on domain-specific terminology, structural patterns, and relevance signals, enabling better performance on domain-specific benchmarks than general-purpose embeddings. Integrates seamlessly with the same vector database infrastructure as general-purpose models.

Offers fine-tuned embedding models tailored to individual company vocabularies, document structures, and relevance signals through a sales-driven engagement process. Custom models are trained on customer-provided data to optimize for company-specific retrieval tasks, terminology, and domain nuances. Requires direct contact with Voyage AI sales team for pricing, timeline, and technical specifications.

Provides voyage-multimodal-3.5, an embedding model that processes both text and images into a shared vector space, enabling cross-modal retrieval (search images with text queries and vice versa). The model is trained on aligned text-image pairs to learn joint semantic representations. Announced but not yet generally available—specific capabilities, context window, and vector dimensionality unknown.

Provides voyage-context-3, an embedding model that generates both chunk-level embeddings (for individual passages) and global document-level context embeddings, enabling improved retrieval accuracy for long documents. The model learns to represent both local semantic meaning and broader document context, reducing false positives in retrieval by understanding how chunks relate to overall document themes. Useful for RAG systems where chunk-level retrieval alone produces irrelevant results.

Provides asynchronous batch processing for embedding large volumes of documents without real-time latency constraints. Batch API is optimized for throughput and cost efficiency, processing documents in bulk and returning results via webhook or polling. Designed for ETL pipelines, data indexing, and periodic re-embedding of large corpora. Technical details (request format, batch size limits, processing time, pricing) not documented.

Provides a dedicated reranking model (voyage-rerank-2.5) that re-orders retrieved results from semantic search to improve relevance and reduce false positives. The reranker takes a query and a list of candidate documents, scoring each document's relevance to the query, and returns a ranked list. Useful as a second-stage ranker in RAG pipelines to improve precision after initial embedding-based retrieval. Operates independently of embedding models.

Provides voyage-rerank-2.5-lite, a smaller variant of the reranking model optimized for inference speed and cost efficiency. Designed for high-throughput ranking scenarios where latency and computational cost are primary constraints. Maintains ranking quality while reducing per-request inference cost compared to voyage-rerank-2.5.

Voyage embeddings are designed to integrate seamlessly with any vector database (Pinecone, Weaviate, Milvus, Qdrant, Chroma, etc.) without custom adapters or preprocessing. Embeddings are returned as standard float32 vectors compatible with all major vector database APIs. No vendor lock-in or proprietary storage format—customers can switch vector databases or embedding providers without data migration.