Voyage AI Freemium

Voyage AI is an embeddings company. It sells text, code and multimodal embedding models plus reranking models that sit in front of a vector store to reorder results by relevance. There is a free token allowance, then pay-per-token pricing.

Why buy embeddings instead of using OpenAI’s?

Because retrieval quality is measurable. Voyage publishes benchmark numbers on MTEB and on domain-specific evals, and its voyage-3 models typically score above general-purpose alternatives on retrieval tasks. For a RAG system, a few points of recall translate into noticeably fewer hallucinated answers.

The company was founded by former Stanford NLP researchers and, notably, was acquired by MongoDB in 2025 to power Atlas Vector Search. It offers an API and does not require you to change your vector database.

Product Features

• voyage-3 family of general text embeddings at 1024 dimensions, with lite and large variants.
• voyage-code-3 trained on source code for code search and repo Q&A.
• voyage-multimodal-3 that embeds text and images into one shared space, useful for PDFs with charts.
• Rerankers (rerank-2 family) that rescore a candidate list and cut irrelevant chunks.
• Domain models for finance and law, trained on sector text.
• Contextualized chunk embeddings that condition each chunk on its document context.

Product Highlights

Two things stand out. First, the reranker. Bolting rerank-2 onto an existing pipeline is a low-effort way to lift answer quality, and in practice a reranker often beats swapping the embedding model. Second, the multimodal model handles scanned documents with figures better than text-only embeddings, which matters for anyone indexing reports.

The trade-off is lock-in and cost. You are paying per token on top of your vector store bill, and switching embedding models later means re-embedding your entire corpus — an expensive operation at scale. If you embed a 50-million-chunk corpus, that migration is a project, not a config change.

Pricing starts with free tokens for evaluation and then per-million-token rates that vary by model; the lite models cost a fraction of the large ones with a modest quality drop.

Use Cases

• RAG over internal documentation where recall matters more than speed.
• Code search across a large monorepo.
• Legal or financial document retrieval using domain models.
• Indexing PDFs with images via multimodal embeddings.
• Adding a reranker stage to an existing search stack.

Quick Summary

• Voyage AI sells embedding and reranking models, not a full RAG platform.
• Its voyage-3 and voyage-code-3 models target retrieval and code search specifically.
• rerank-2 rerankers often improve relevance more cheaply than re-embedding.
• Free evaluation tokens, then per-token billing layered on top of your vector DB.
• MongoDB acquired Voyage AI in 2025, tying it to Atlas Vector Search.