Voyage AI

What is Voyage AI?

Voyage AI is a model and API platform for semantic search, retrieval-augmented generation, and unstructured-data discovery. Embedding endpoints convert text and other supported data into vectors for a vector database, while rerankers score query-document relevance and reorder an initial set of candidates. Better retrieval can give downstream language models more relevant context and reduce wasted tokens, but it does not guarantee factual answers.

The model portfolio includes general-purpose, lightweight, contextual, multimodal, multilingual, code, finance, and legal options, plus company-specific models and custom deployment. Models can be accessed through the Voyage API, cloud and data-platform offerings, in-VPC arrangements, or custom and on-premises licensing. Batch jobs provide an asynchronous option for large corpora and current documentation advertises a 33% discount compared with standard inference.

How to Use Voyage AI

1. Choose a model based on language, domain, modality, context, dimensionality, and latency.
2. Normalize and chunk the corpus, preserving document identifiers and access controls.
3. Generate embeddings and store them in a compatible vector database.
4. Embed each query, retrieve candidate documents, and apply a reranker when useful.
5. Pass only authorized, high-scoring context to the language model with citations.
6. Evaluate recall, ranking quality, latency, cost, bias, and access leakage on real queries.

Core Features

• Text embeddings: Generates semantic vector representations for documents and queries.
• Contextual embeddings: Preserves broader document context while representing chunks.
• Multimodal retrieval: Supports search across text and supported visual inputs.
• Rerankers: Reorders retrieved candidates using detailed query-document relevance.
• Domain models: Offers specialized retrieval for code, finance, legal, and multilingual data.
• Deployment flexibility: Provides API, cloud, in-VPC, custom, licensed, and on-premises paths.

Use Cases

• Build semantic search over product, support, or enterprise documents.
• Improve RAG context selection before answer generation.
• Retrieve relevant code, legal materials, or finance documents.
• Reorder lexical or vector search results with instruction-aware relevance.

Pricing

Pricing is usage-based by processed tokens. Current text models range from $0.02 to $0.18 per million tokens depending on model, and listed rerankers range from $0.02 to $0.05 per million processed tokens. Many current models include the first 200 million tokens free per account, while selected older domain models include 50 million. Multimodal images use pixel-based token equivalents. Confirm the current model table before estimating production cost.

Frequently Asked Questions

Are embeddings the same as generated answers?

No. Embeddings represent semantic meaning for retrieval; another system usually generates or displays the final response.

Must Voyage rerankers be used with Voyage embeddings?

No. Rerankers can refine candidates from other embedding or lexical search methods.

Does stronger retrieval eliminate hallucinations?

No. It improves available context, but applications still require citations, prompt controls, access filtering, and answer evaluation.