Embeddings

Voyage AI

Voyage AI’s embedding models, including domain models for code, finance and law. Voyage’s own API lives at /v1/embeddings in a shape close to OpenAI’s, so VectorAxis serves both on that one endpoint: the Voyage AI SDK and the OpenAI SDK work against it. Use a Voyage AI virtual key (Bring your own key, with your Voyage API key). Voyage AI keys serve embeddings only; they are refused on chat completions.

With the Voyage AI SDK

Python

import voyageai

vo = voyageai.Client(
    api_key="vk-1a2b3c4d5e6f7a8b.Xk3pQ9vN2mR7tY1wZ8bC4dF6gH0jK5lM3nP7qS9uV2w",  # a Voyage AI virtual key
    base_url="https://api.vectoraxis.ai/v1",                                       # the SDK appends /embeddings
)

result = vo.embed(["What is our refund policy?"], model="voyage-3.5", input_type="query")
print(len(result.embeddings[0]))

Voyage’s request fields reach Voyage as you send them: input_type (query or document), truncation, output_dimension, output_dtype and encoding_format. See Voyage’s API reference.

With the OpenAI SDK

Python

client.embeddings.create(
    model="voyage-3.5",
    input=["Refunds are issued within 14 days."],
    dimensions=512,
    extra_body={"input_type": "document"},
)

OpenAI field	Sent to Voyage as
dimensions	output_dimension
user	dropped (Voyage has no such field)
every other field	as sent

Voyage’s answer already has OpenAI’s shape; VectorAxis adds usage.prompt_tokens next to Voyage’s usage.total_tokens, which OpenAI clients read.

Cost and logs

Each call is priced on the tokens Voyage reports, at the model’s rate, and logged with operation embedding. Budgets, guardrails, retries and the exact-match cache (x-cache-mode: simple) work as on every embeddings call. Voyage’s rerank, multimodal and contextualized-chunk endpoints are not available through VectorAxis yet.

VectorAxis is built by GK Labs, a member of NVIDIA Inception.