API reference
Embeddings
Dashboard
BetaWe just launched Apps: hosted models, actions and pages, on the same bill as your inference.Read the Apps docs

Embeddings

POSThttps://api.inferencebureau.com/v1/embeddings

Turn text into vectors, in the OpenAI embeddings shape. Each request is admitted, held and settled like a chat request, and billed on input tokens only.

Body

application/json, up to 1 MiB; unknown fields refused
  • modelstringrequired

    A certified embedding model, such as openai/text-embedding-3-small. GET /v1/models lists the ones your key may use with type: "embedding". A vendor's own name (text-embedding-3-small) works too, as does a team alias that names an embedding model. A chat model is refused with 400 invalid_request.

  • inputstring | string[] | number[] | number[][]required

    One text, up to 2,048 texts, or the same as token ids. Texts can't be empty. The whole request is limited to 1 MiB.

  • dimensionsintegeroptional

    Shorter vectors, on models that support it: 1 to 8,192. Passed to the provider as given.

  • encoding_formatfloat | base64optional

    Defaults to float.

  • userstringoptional

    Your end user, recorded for attribution and per-end-user Directives, and not forwarded.

  • ibobjectoptional

    tags (up to ten) and session, as on chat completions. Not forwarded.

  • providerOptionsoptional

    Accepted for compatibility with AI SDK clients, and not forwarded.

Headers

  • Authorizationstringrequired

    Bearer ib_live_…, or the key in x-api-key. See Authentication.

  • X-Client-Request-IDstringoptional

    Your own correlation id, recorded on the receipt.

  • ib-sessionstringoptional

    Groups the request with others in Logs, such as the batch it belongs to.

  • ib-userstringoptional

    Your end user, up to 256 characters; wins over the body's. See End users and tags.

  • ib-tagsstringoptional

    Comma-separated tags for cost breakdowns, at most ten with ib.tags. Never sent upstream.

Notes

  • Failoveroptional

    If a provider fails before answering, the request is tried once on another certified provider of the same model. It never moves to a different model, because vectors from different models cannot be compared.

  • Costoptional

    The hold assumes one token per two characters, and settles at what the provider billed. Look up the receipt with GET /v1/requests/:id using the X-Request-ID response header.

  • Directivesoptional

    Model and provider allowlists, budgets, rate limits, maxInputTokens, maxRequestChargeUsd and requireZdr apply. A team that requires zero data retention is routed only to providers that offer it. The estimated input counts against tokensPerMinute.

  • Limitsoptional

    A call has 60 seconds. Responses are not streamed, and Idempotency-Key is not supported here.