Introducing the embeddings API

The inference API now supports embeddings. Send a POST /v1/embeddings request with a model from the catalogue and get back a list of floating-point vectors. The endpoint is OpenAI-compatible, so existing code that calls /v1/embeddings works by changing the base URL and the key.

Embeddings are how you feed semantic meaning into downstream systems: similarity search, document clustering, reranking retrieved passages before they reach the model, and anything else that needs to compare texts numerically. They complete the inference surface the platform offers, governed the same way as chat completions.

EU-sovereign by default

The default model is mistral/mistral-embed, served in the EU. Text you send to it never leaves European infrastructure. No external-routing acceptance is needed; the request goes straight through.

curl https://api.akumi.cloud/v1/embeddings \
  -H "Authorization: Bearer wk_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral/mistral-embed",
    "input": "The quarterly report is ready for review."
  }'

Pass input as an array to embed a batch of texts in one call. The response contains one embedding object per string, in order, each with its own index.

External models, with the guard

openai/text-embedding-3-small and openai/text-embedding-3-large are available for organisations that have recorded the non-EU-routing acceptance. This is the same acceptance used for external chat models, enforced by the same egress guard. Without it, the request is rejected. With it, the full embedding catalogue is open.

No firewall, by design

The PII firewall pseudonymises text before it reaches a model and restores it in the generated output. That pattern does not translate to embeddings: pseudonymising the input embeds the wrong text, there is no output text to restore into, and the per-request token map means the same entity maps to a different placeholder across calls, making vectors non-comparable. Embeddings are therefore governed by residency, not the firewall. The EU model stays in the EU. External models require the acceptance. The governance is correct for the data type, not applied by habit.

Metered and audited

Embeddings are priced per input token from the model's per-1,000-token rate. There are no output tokens. They appear on the Usage page as a separate line from chat completions. Every request is written to the audit trail with the model, the token count, and the acting key, but never the content of the text you embedded.

Get started

Point any client at the endpoint and switch to mistral/mistral-embed. Your semantic search pipeline runs without any data leaving the EU. The embeddings guide covers the full request shape, the models table, and the dimensions parameter.