Embedding models and rerankers for search and retrieval over unstructured data
Voyage AI is an embedding and reranking API platform for building RAG and semantic search pipelines.
AI Panel Score
6 AI reviews
Reviewed
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.Voyage AI is an embedding and reranking API platform for building RAG (Retrieval-Augmented Generation) and semantic search pipelines. It offers general-purpose embedding models, domain-specific models optimized for finance, legal, and code, and company-specific fine-tuned models, all designed to improve retrieval quality. The models are built around reduced vector dimensionality, which cuts storage and search costs by roughly 3x to 8x versus standard embeddings, and long context windows up to 32K tokens, the longest commercial offering in the category. Pricing is usage-based pay-as-you-go with a free tier and trial, though per-token rates are not published. Other capabilities include multimodal embeddings, a batch API, and plug-and-play vector database and LLM integration. TopReviewed's six-seat AI review panel scored it 8.1/10, praising the MongoDB acquisition for settling vendor-viability questions while noting the lack of public pricing makes cost modeling difficult before integration. It best fits ML engineers building domain-specific RAG pipelines.
Voyage AI fits into a RAG pipeline between unstructured data and an LLM: documents are converted to vector embeddings using Voyage's models, stored in a vector database, and retrieved via similarity search. A reranker then re-scores candidate results before they are passed to the LLM, improving relevance and reducing the chance of irrelevant context reaching the model. Users interact with the system through an API, sending text or multimodal content and receiving embeddings or ranked lists in return.
The platform includes several model families. General-purpose models handle multilingual and cross-domain content out of the box. Domain-specific models—covering finance, legal, and code—are tuned on industry data for higher retrieval accuracy in those verticals. Multimodal models (including voyage-multimodal-3.5) support combined text and image inputs. Rerankers such as rerank-2.5 and rerank-2.5-lite add instruction-following capability and sit at a different price-performance point. A Batch API is available for large-scale, asynchronous workloads. The company also highlights voyage-context-3, which is designed to capture both chunk-level details and broader document context simultaneously.
Voyage AI targets developers and ML engineers building search, question-answering, or document retrieval systems who need embedding and reranking infrastructure without training their own models. Competitors in the embedding API space include OpenAI Embeddings, Cohere Embed, and Google Vertex AI Embeddings. Pricing is usage-based; a free tier appears to be available for initial access, with paid usage billed per token.
The product is API-first and designed to be plug-and-play with any vector database (Pinecone, Weaviate, Chroma, etc.) and any LLM. It is accessed via web API, making it platform-agnostic for backend integration on any operating system or cloud environment.
Provides embedding models highly optimized for industry-specific data including finance, legal, and code use cases.
Generates vector embeddings from unstructured data for use in search and retrieval pipelines, with general-purpose models ready for any purpose and language out-of-the-box.
Supports multimodal retrieval via voyage-multimodal-3.5, enabling embeddings across multiple data modalities beyond text.
Reranks retrieved documents to improve relevance before passing context to an LLM, with models like rerank-2.5 supporting instruction following and a new price-performance frontier.
A specialized model that captures focused chunk-level details while retaining global document context for improved retrieval accuracy.
Provides a Batch API for simpler and more efficient processing of large-scale embedding and retrieval workloads.
Offers 2x cheaper inference costs compared to competing models while maintaining superior retrieval accuracy.
Supports the longest commercial context length available at 32K tokens, allowing retrieval over large documents without chunking limitations.
Produces vectors that are 3x–8x shorter than standard embeddings, reducing vector search and storage costs significantly.
Delivers 4x smaller model size and faster inference speeds while maintaining superior retrieval accuracy compared to larger alternatives.
Offers fine-tuned models trained on a company's unique data and terminology to act as specialized librarians for proprietary content.
Designed for modular integration with any vector database and large language model, enabling flexible RAG pipeline assembly.
Usage-based access to Voyage AI embedding models and rerankers, billed per token consumed
Fine-tuned, company-specific models optimized for your organization's unique data and terminology
MongoDB paid $220M for this — that's your viability answer.
“Voyage AI is a focused embedding and reranking API with real technical differentiation: 32K context windows, domain-specific models, and costs that undercut OpenAI Embeddings. The MongoDB acquisition in February 2025 eliminates most survival risk.”
MongoDB acquired Voyage AI for $220 million in February 2025. That's not a feature — that's a balance sheet backstop. Vendor viability concern gone.
The technical case is specific. Voyage-context-3 captures chunk-level and document-level context simultaneously, which solves a real RAG failure mode. Low-dimensionality vectors run 3x–8x shorter than standard embeddings, cutting storage and search costs. Rates are posted per token — $0.02 to $0.12 per million on the current models, first 200 million free — so you can model cost before you commit.
Two things make this a strong bet. One: domain-specific models for finance, legal, and code mean the embeddings actually understand your content, not just tokenize it. Two: plug-and-play with Pinecone, Weaviate, or any vector DB means no lock-in beyond the model layer itself. Pilot it on a single RAG workload and compare retrieval quality against Cohere Embed head-to-head — every benchmark they publish is their own.
Outpaces OpenAI Embeddings on context length (32K tokens) and cost, and publishes NDCG@10 comparisons against Cohere Embed directly — though every benchmark is run in-house.
A $220M MongoDB acquisition makes this a board-defensible choice; no reputational downside visible from the outside.
API-first, plug-and-play with any vector DB, and a free tier means a working integration in days, not quarters.
Domain-specific models for finance, legal, and code advance retrieval quality, not just cost reduction on existing pipelines.
Acquired by MongoDB for $220M in February 2025 — runway and survival are no longer the question.
ML engineers building RAG pipelines in finance, legal, or code who need domain-tuned embeddings without training their own models.
Your workloads are simple keyword search and you don't have a vector database already in the stack.
MongoDB's $220M bet on embeddings gives Voyage real infrastructure credibility.
“Voyage AI is a focused embedding and reranking API with genuine craft depth — 32K context windows, low-dimensionality vectors, domain-specific models for finance, legal, and code. The MongoDB acquisition changes the risk calculus significantly for anyone evaluating long-term vendor stability.”
The model architecture tells you something important: reduced vector dimensionality (3x–8x shorter than standard) plus 32K token context windows isn't a marketing slide — it's a deliberate infrastructure position. Someone made real tradeoffs to get there. voyage-context-3 capturing both chunk-level and document-level context simultaneously is the kind of retrieval nuance that separates teams who've actually debugged RAG precision from teams who haven't.
The MongoDB acquisition at $220M is the biggest strategic signal here. If you're building on Voyage today, in 3 years you're almost certainly building on a feature inside Atlas. That's either a moat or a migration risk depending on your stack. Teams already on MongoDB benefit enormously; teams on Pinecone or Weaviate should watch how the integration surface evolves post-acquisition.
Against Cohere Embed and OpenAI Embeddings, Voyage's domain-specific vertical models are a genuine differentiator for finance or legal retrieval workloads. Pricing is posted per token — $0.02 to $0.12 per million — though company-specific fine-tuning stays quote-only.
Voyage sits above OpenAI Embeddings on retrieval craft and above generic APIs on vertical depth, but post-acquisition positioning vs. Cohere is still settling.
Domain-specific models for finance, legal, and code plus HIPAA and SOC 2 compliance map directly to how ML engineers build production RAG pipelines.
Plug-and-play with any vector DB and LLM, Batch API for async workloads — the integration surface is intentionally non-opinionated and stack-agnostic.
MongoDB acquisition de-risks vendor survival but introduces potential Atlas lock-in pressure over a 3-year horizon.
Low-dimensionality vectors, 32K context, instruction-following rerankers like rerank-2.5 — this is library-grade embedding infrastructure, not a thin API wrapper.
ML engineers building production RAG pipelines in finance, legal, or code domains who need best-in-class retrieval without training their own models.
Your architecture is fully committed to a competing vector DB ecosystem that MongoDB may eventually treat as a second-class integration target.
Published rate card: $0.02–$0.12 per million tokens, first 200M free, no sales call.
“Voyage AI publishes a full per-token rate card — $0.02 to $0.12 per million, first 200 million free — so TCO modeling needs no sales call. MongoDB acquired them for $220M in February 2025, and pricing strategy could still shift.”
Rates are published per token. $0.02 to $0.12 per million, depending on model. First 200 million tokens are free. No seats, no minimum, no platform fee. Procurement won't fight this one.
The efficiency numbers are real too. Low-dimensionality vectors run 3x–8x shorter than standard embeddings — storage costs drop materially. 32K token context means fewer chunking workarounds. Run 1B tokens a month through voyage-4: $60. That's $720 a year. The embedding line is rounding error next to the vector DB bill.
The MongoDB acquisition adds a variable. Enterprise pricing could tighten post-integration. Throughput is tied to cumulative spend — $100 billed unlocks 2x rate limits, $1,000 unlocks 3x. SOC 2 and HIPAA compliance are confirmed — one less procurement blocker. Fine-tuned models still require a sales conversation.
SOC 2 and HIPAA confirmed, self-serve signup, no seat licence and no minimum — procurement has very little left to negotiate.
Pay-as-you-go with no term commitment and no renewal to escape; the catch is throughput scaling with cumulative spend, and fine-tuning staying sales-gated.
Full per-model rate card published without a login — $0.02 to $0.12 per million tokens, first 200 million free — and only the fine-tuned tier is quote-only.
Retrieval quality improvements are measurable via benchmark comparisons; 2x cheaper inference claim gives a quantifiable starting point.
3x–8x smaller vectors cut storage costs, and published per-token rates make year-3 TCO a spreadsheet exercise rather than a guess.
ML engineers building RAG pipelines in finance, legal, or code who want domain-specific embeddings on published per-token rates.
Your finance team needs a fixed annual commitment rather than billing that moves with token volume.
Voyage AI: Purpose-Built Embedding API That Earns Its Spot in RAG Pipelines
“32K context window and domain-specific models for finance, legal, and code make this a serious upgrade over OpenAI Embeddings for specialized retrieval. The MongoDB acquisition in February 2025 for $220M signals durability, and per-token rates are published openly — $0.02 to $0.12 per million.”
The API-first design is the right call. Plug-and-play with Pinecone, Weaviate, Chroma — no adapter layers, no vendor lock-in on the vector DB side. The Batch API for async workloads tells me someone built this after watching teams hammer synchronous endpoints during document ingestion. voyage-context-3 capturing both chunk-level and document-level context simultaneously is the kind of detail that matters at 3am when retrieval quality is tanking a demo.
The 32K token context window is the clearest differentiator versus Cohere Embed or OpenAI's text-embedding-3 series. That's not a marketing number — that's fewer chunking headaches in production. Low-dimensionality vectors at 3x–8x shorter than standard means real storage savings at scale. The rerank-2.5 instruction-following capability is a workflow upgrade for pipelines where relevance criteria shift by query type.
Pricing is the pleasant surprise. Per-token rates are posted — voyage-4 at $0.06 per million, first 200 million free — so cost modeling is a spreadsheet, not a sales email. Company-specific fine-tuned models still need a sales contact, which is expected at that tier. The real day-3 friction is generation churn: new models land every few months and pinning a version is on you.
API-first with broad vector DB compatibility makes integration straightforward, and published per-token rates mean cost modeling is arithmetic once the free 200M tokens run out.
Quick Start Tutorial, a public changelog, and a homepage 'Try it now' path — the docs are built for someone shipping, not someone evaluating.
Remaining friction is generation churn and sales-gated fine-tuning; rates and release notes are both public, so neither costs you a support ticket.
Domain-specific models, instruction-following rerankers, voyage-context-3's dual-context architecture, and company-specific fine-tuning give clear depth progression from beginner to production-scale usage.
Modular RAG pipeline design with Batch API and plug-and-play integrations fits naturally into standard ML engineering workflows without demanding new tooling habits.
ML engineers building domain-specific RAG pipelines in finance, legal, or code who need best-in-class retrieval without training their own embedding models.
You need a fixed-price plan rather than billing that moves with token volume.
Embedded deep in MongoDB now, and the retrieval quality story is real
“Voyage AI is a serious embedding and reranking API for developers who need better RAG pipelines without training their own models. The $220M MongoDB acquisition in February 2025 either cements its future or complicates its independence — probably both.”
The 32K token context window is the headline number here. Most embedding APIs force you to chunk aggressively and lose document-level coherence. Voyage's voyage-context-3 model is specifically built to hold chunk-level detail and broader document context at once — that's not marketing fluff, that's a real pipeline problem being solved. Domain-specific models for finance, legal, and code mean you're not forcing a general-purpose model to learn your terminology on the fly.
The tradeoff is that this is deeply a developer product. No UI to speak of. No dashboard polish to review. Daily Polish and Mobile Parity scores here are almost irrelevant — this lives in your backend, not your browser. Compared to OpenAI Embeddings, Voyage is a deliberate specialist play: narrower surface area, higher retrieval accuracy claim, lower vector dimensionality reducing storage costs 3x–8x.
Pricing is usage-based with a free tier, which is the right call for an API product. The rates are posted per model too — two cents to twelve cents per million tokens, first 200 million free, no 'contact us' wall. The MongoDB acquisition adds long-term stability questions — great infrastructure backer, but enterprise absorption can slow the roadmap fast.
API-first product with docs, a quick-start tutorial, and a public changelog — polish lives in the DX rather than a UI, and the DX holds up.
Plug-and-play with Pinecone, Weaviate, Chroma, and any LLM keeps onboarding fast, but fine-tuned company-specific models require a sales conversation to unlock.
Web API product — mobile parity is essentially irrelevant, and the score reflects that this dimension simply doesn't apply here.
Quick Start Tutorial in docs plus a 'Try it now' homepage CTA suggests a low-friction first ten minutes for any developer who's integrated an API before.
A $220M acquisition target and HIPAA plus SOC 2 compliance signals production-grade reliability expectations; Batch API for async large-scale workloads reinforces this.
ML engineers and backend developers building RAG pipelines who need production-grade retrieval without training their own embedding models.
You want a self-serve, no-code search experience rather than an API you wire up yourself.
MongoDB's $220M acquisition either validates this or complicates it — probably both
“Solid embedding API with real technical differentiation: 32K context, low-dim vectors, domain-specific models for finance/legal/code. The MongoDB acquisition in February 2025 is the single biggest variable to watch.”
Three green flags first. The 32K token context window beats OpenAI Embeddings and Cohere Embed on docs I can find. Low-dimensionality vectors — 3x–8x shorter — aren't marketing padding; smaller indexes mean real cost savings at scale. Domain-specific models for finance and legal is a legitimate moat. Most competitors don't bother.
One real yellow flag. MongoDB acquired this for $220M in February 2025. Maybe they leave it standalone. Maybe it becomes Atlas-only. Category history says acquisitions frequently collapse standalone access within 18–24 months. We're past that window and the API is still open — but the branding now reads 'Voyage AI by MongoDB.' Watch the terms.
Exit portability is actually decent. Embeddings are just vectors — swap to Cohere or OpenAI Embeddings, re-embed your corpus, done. Painful but not catastrophic. The reranker dependency (rerank-2.5) is stickier. If this goes MongoDB-exclusive, that's the real migration cost.
32K context window, low-dim vectors, and domain-specific models (finance, legal, code) are concrete gaps vs. OpenAI Embeddings — not just feature-list fluff.
Embeddings re-generation is feasible with OpenAI or Cohere as fallback; reranker dependency is the harder lock-in risk if MongoDB closes the API.
A public changelog and model releases every few months show real shipping cadence; MongoDB's endgame for the standalone API is the open question, not the pace.
Claims like '2x cheaper' and '4x smaller model size' are specific and falsifiable, and published per-token rates let you check the cost story — the benchmarks behind them are all self-run.
Matches patterns of specialist API players that carved real niches (like Cohere early days), not vaporware — but acquisition-stage companies have a mixed survival record as standalone products.
ML engineers building RAG pipelines over finance, legal, or code documents who need better retrieval than OpenAI Embeddings without training their own models.
You're worried about API continuity if MongoDB eventually folds this into Atlas exclusivity.
Common questions answered by our AI research team
Yes, Voyage AI is HIPAA compliant. HIPAA is listed alongside SOC 2 under Privacy and Compliance.
Voyage AI models support up to 32K tokens, the longest commercial context length available.
Yes, Voyage AI offers domain-specific models highly optimized for finance, legal, and code.
Yes, Voyage AI is plug-and-play with any vector database and LLM.
Get started via the Quick Start Tutorial in the Docs, or click "Try it now" on the homepage to begin using embeddings directly.
Who Voyage AI competes with, and how each one differs
Cohere ships its own Embed and Rerank model lines via API, competing directly in both embeddings and reranking; Voyage's own January 2025 benchmarks reported voyage-3-large beating Cohere-v3-English by ~20.7% and rerank-1 beating Cohere's rerank-english-v3/rerank-multilingual-v3.
OpenAI's text-embedding-3 models are the default embeddings choice for developers already on the OpenAI API; Voyage's January 2025 benchmarks claimed voyage-3-large outperforms OpenAI-v3-large by ~9.7% on retrieval tasks across 100 datasets.
Jina AI ships open-weight embedding models (jina-embeddings-v5, Apache-2.0/CC-BY-NC licensed) deployable outside a hosted API, competing with Voyage on the self-hostable segment Voyage's proprietary models don't serve.
Company
Voyage AIFounded
2023Pricing
Usage-basedFree Trial
AvailableFree Plan
AvailableVoyage AI is a Palo Alto, California-based AI research company developing embedding and reranking models for retrieval-augmented generation, acquired by MongoDB in February 2025.
