Every retrieval-augmented generation (RAG) pipeline needs two models: an embedding model that converts text into vectors so similar passages can be found by nearest-neighbor search, and a generation model — Claude — that writes the answer. Anthropic's position is stated plainly in its docs: "Anthropic does not offer its own embedding model," and it names Voyage AI as one embeddings provider "that has a wide variety of options and capabilities." The docs also hedge — "you should assess a variety of embeddings vendors to find the best fit for your specific use case" — and so should you. What follows is how the documented Voyage lineup fits a Claude pipeline.
The general-purpose tier: the Voyage 4 family
The current recommended generation is the Voyage 4 family: voyage-4-large, voyage-4, voyage-4-lite, and voyage-4-nano (the nano model is open-weight under Apache 2.0). All four share a 32,000-token context window and a 1024-dimension default, with 256, 512, and 2048-dimension options. The family is a quality/cost ladder: start with voyage-4, step up to -large if retrieval quality is your bottleneck, step down to -lite or -nano if embedding cost or self-hosting matters more.
Domain-specific and contextualized models
Voyage also publishes specialists, and the docs list them by domain: voyage-code-3 for code, voyage-finance-2 for finance, and voyage-law-2 for legal (16K context). There is a multimodal model, voyage-multimodal-3.5, covering text, images, and video.
The most interesting entries for RAG architects are the contextualized chunk models, voyage-context-4 and voyage-context-3 (120K context, called via contextualized_embed()). Normal embedding treats each chunk in isolation, so "the termination clause above notwithstanding" embeds poorly because the model cannot see what "above" refers to. Contextualized models embed chunks with awareness of surrounding document context, attacking that problem at the index level. See contextualized embeddings for the deeper dive. Rounding out the lineup are the rerankers rerank-2.5 and rerank-2.5-lite, which re-score your top-k candidates — covered in the reranking pattern.
The setting everyone forgets: input_type
Voyage models want to know whether a text is a query or a document, because it prepends different retrieval prompts to each. The docs are unambiguous: always set input_type="query" when embedding a search query and input_type="document" when embedding corpus text. Skipping it (or using the same value for both sides) quietly degrades retrieval quality — the kind of bug no error message will ever surface.
A helpful property for your vector database: Voyage embeddings are normalized to length 1, so dot-product and cosine similarity give identical rankings. You can pick whichever your store computes faster.
Shrinking the index: quantization and Matryoshka truncation
Vector indexes get expensive at scale — millions of 1024-dimension float vectors add up. Current Voyage models support two documented levers:
- Quantization via
output_dtype:float,int8/uint8, orbinary/ubinary. Lower-precision types store each dimension in fewer bits, trading a little accuracy for a much smaller, faster index. - Matryoshka dimension truncation: the models are trained so that the first N dimensions of a vector are themselves a usable smaller embedding, letting you truncate 1024-dimension vectors down (e.g., to 256) without re-embedding.
Where it plugs into Claude on 3P platforms
Voyage is a separate vendor with its own API and pricing — Anthropic's docs deliberately do not list Voyage prices and link to Voyage's own pricing page. For AWS-centric stacks, the docs note that Voyage embeddings are also available through AWS Marketplace, which keeps billing inside your AWS relationship. One boundary to be clear about: Amazon Bedrock Knowledge Bases documents only Amazon Titan and Cohere embedding models for customer-managed knowledge bases — Voyage is not in that list — so a Voyage-based index means running your own retrieval and passing results to Claude via search_result content blocks, which works on all four platforms. Compare with Bedrock KB's Titan vs Cohere choice if you would rather stay fully managed on AWS.
Where to go next
Start from RAG basics, then design chunking with the chunking pattern and retrieval quality with hybrid retrieval.