What Voyage AI does
Voyage AI provides hosted embedding and reranking APIs for general, code, multilingual, finance, legal, contextual, and multimodal retrieval workloads.
Voyage AI concentrates on the retrieval layer rather than generating final prose. Its catalog includes general-purpose, contextual, code, multilingual, legal, finance, and multimodal embedding models plus rerankers. This specialization lets teams compare domain and storage tradeoffs without replacing the rest of their RAG stack.
The current pricing documentation grants substantial free token allowances for several model families, then meters text by tokens, multimodal text and pixels, and reranking by processed tokens. Rates and allowances differ by model. Storage, vector database operations, image preprocessing, video frames, re-indexing, and generator calls remain separate costs; confirm eligibility and current prices before planning a production budget.
Vendor benchmark gains do not guarantee improvement on a private corpus. Chunking, labels, query mix, access metadata, duplicates, and the vector database can dominate performance. Domain models may encode specialist assumptions or uneven coverage. Build a reviewed evaluation set, prevent test leakage, protect embeddings and documents, preserve citations, and require abstention or review when retrieval is weak.
How Voyage AI works
A client sends documents and queries to an embedding endpoint with the proper input type, receiving normalized numeric vectors for storage and nearest-neighbor search. A reranking endpoint can then score retrieved query-document pairs, and multimodal models embed text, images, or sampled video frames into a shared space. The application chooses thresholds, retains source metadata, and sends only approved evidence to a generator. Humans validate retrieval and resulting claims.
How to set up Voyage AI
Create a domain evaluation set
Collect representative queries, relevant and hard-negative passages, languages, modalities, ACLs, and high-cost failure cases.
Compare model and dimension choices
Test general, domain, code, contextual, or multimodal models for recall, precision, storage, latency, and upgrade compatibility.
Secure ingestion
Keep keys server-side, minimize submitted data, preserve tenant permissions and source versions, and document retention and deletion.
Add reranking selectively
Retrieve a broad candidate set, rerank only enough passages to improve precision, and measure token and latency cost against lift.
Operate versioned indexes
Record model IDs and input types, stage re-indexing, monitor drift, show citations, and require human review for consequential outputs.
Voyage AI FAQs
Does Voyage AI generate answers?
Its core APIs produce embeddings and relevance scores; a separate application or language model generates the final answer.
Is there a free allowance?
Yes. Current documentation lists model-specific free token and pixel allowances before usage charges begin.
What is reranking?
Reranking applies a more expensive relevance model to a smaller retrieved set to improve ordering and precision.
Can Voyage embed images and video?
Supported multimodal models embed text, images, and sampled video frames, with pixel-based usage metering.
Will a domain model always perform better?
No. Evaluate it on representative, independently labeled data; corpus structure and query mix may favor another model.
Listing reviewed 2026-07-30. Product details and pricing can change; verify important terms on the provider's website.
Related Coding AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.