Voyage AI

✓ Editorially verified

State-of-the-art embedding models and rerankers purpose-built for retrieval and RAG.

Freemium· Free tier: 200M free text tokens per account for current models (50M for older specialized). Text embeddings $0.00002–$0.00018 per 1K tokens depending on model tier. Rerankers $0.00002–$0.00005 per 1K tokens after 200M free. Multimodal $0.12 per 1M text tokens + $0.60 per 1B pixels. Batch API 33% discount. File storage $0.05/GB/month.RAGin-house (voyage-3.5, voyage-4 series, voyage-code-3, voyage-finance-2, voyage-law-2, voyage-multimodal-3.5, voyage-context-3, rerank-2.5)

Visit website →

Best for

RAG and semantic-search teams who care about retrieval quality, need domain-tuned embeddings (code, finance, legal), and want a drop-in API that plugs into their existing vector database and LLM stack.

Skip if

Teams that need on-prem or self-hosted embedding models, want an all-in-one RAG platform with a vector DB bundled in, or have strict data-residency requirements that forbid third-party APIs.

Voyage AI (now part of MongoDB) builds embedding models and rerankers that consistently top public retrieval benchmarks such as MTEB, and it has become one of the default choices when teams want higher recall than OpenAI's text-embedding-3 or open-source alternatives can deliver. The platform is deliberately narrow: it does not ship an LLM or a vector database, only the neural components that turn documents and queries into vectors and then re-rank the top candidates. That focus is the point — Voyage's models routinely beat larger general-purpose embeddings while producing shorter vectors (often 3x–8x smaller), which reduces storage costs and speeds up nearest-neighbor search in whichever vector store you already run.

The catalog covers general-purpose models (voyage-3.5, voyage-4-large, voyage-4-lite), domain-specialised variants for code, finance, and legal, a long-context 32K model, a multimodal model that embeds text and images into a shared space, voyage-context-3 which produces chunk-level vectors informed by surrounding document context, and the rerank-2.5 family with instruction-following. Everything is exposed through a simple REST API and official Python/TypeScript SDKs, and there are first-class integrations with LangChain, LlamaIndex, Haystack, Pinecone, MongoDB Atlas, Weaviate, Qdrant, and every other major vector DB. Databricks, Anthropic, Replit, Snowflake, and AWS are named customers.

Typical workflows: powering a production RAG chatbot over a proprietary knowledge base, semantic search across legal contracts or SEC filings using the domain models, code search inside a monorepo with voyage-code-3, and two-stage retrieval where a cheap ANN pass over embeddings is refined by rerank-2.5 before results hit the LLM. A generous 200M-token free tier per account makes it realistic to prototype and even run small production workloads without a paid contract.

Editor's take

If retrieval quality is what's holding your RAG app back, Voyage is the shortest path to a measurable improvement — swap the embedding endpoint, keep everything else, and re-run your eval. The domain models and rerankers are genuinely better than defaults, not marketing. My main hesitation is strategic: post-MongoDB, standalone Voyage may increasingly optimise for Atlas users first.

— The AI Tool Bible editorial team

Pros

✅ Consistently near the top of MTEB and BEIR retrieval leaderboards — measurable recall gains over OpenAI text-embedding-3-large in most public evaluations.
✅ Short output dimensions (as low as 256 or 512) cut vector storage and ANN latency 3x–8x versus 1536/3072-dim competitors.
✅ Domain-tuned models (code, finance, legal) meaningfully outperform general embeddings on in-domain corpora.
✅ voyage-context-3 embeds chunks with awareness of surrounding document context, reducing the classic 'lost context' problem in fixed-window chunking.
✅ Rerank-2.5 with instruction-following gives a clean two-stage retrieval pipeline without training a custom cross-encoder.
✅ Generous 200M-token free tier per account makes prototyping and small production workloads essentially free.
✅ Batch API offers a 33% discount for large offline embedding jobs.
✅ MongoDB acquisition (2025) means tight, ongoing integration with Atlas Vector Search.

Cons

⚠️ API-only closed models — no self-hosting option, so latency-sensitive or air-gapped deployments are ruled out.
⚠️ Not an end-to-end RAG stack — you still need a vector database, LLM, and orchestration layer, which increases integration surface.
⚠️ Post-MongoDB acquisition, product roadmap and standalone longevity depend on MongoDB's priorities.
⚠️ Domain models cover finance, legal, and code but nothing else — medical, scientific, or multilingual-heavy corpora fall back to general models.
⚠️ Documentation is competent but sparser than OpenAI's or Cohere's — fewer end-to-end recipes for advanced patterns like hybrid search or query expansion.
⚠️ Pricing per token is competitive but not the cheapest — self-hosted open models (e.g. BGE, E5) are free at inference if you have GPUs.

Use cases

Production RAG chatbot over proprietary docsTwo-stage retrieval with embed + rerankCode search across a monorepoLegal contract semantic searchFinancial filings and research retrievalMultimodal image-and-text searchLong-context document embedding (32K tokens)Context-aware chunk embedding for dense passagesBatch embedding of large historical corporaMongoDB Atlas Vector Search backends