Embeddings

POST /v1/embeddings turns text into vectors in OpenAI's Embeddings format. Point the OpenAI SDKs, LangChain or LlamaIndex at https://api.zdr.sh/v1 and they work. The request is served by a model the catalog marks as an embedding model; Models lists the ones your key can use.

Request#

Field	Type	Notes
`model`	string	Required. An embedding model id. A chat model is refused with `400 model_endpoint_mismatch`
`input`	string or array	Required. The text, or up to 2,048 texts. No item may be empty, each must fit the model's window, and the body is at most 4 MiB
`encoding_format`	string	`float` (default) or `base64`
`dimensions`	integer	The vector's size, within the model's range
`user`	string	An end-user id recorded with the usage. It is never sent to a model

An array of token ids is refused: another tokenizer's vocabulary would give a wrong vector without an error.

Response#

json

{  "object": "list",  "data": [{ "object": "embedding", "index": 0, "embedding": [0.0023, -0.0091] }],  "model": "<your model id>",  "usage": { "prompt_tokens": 6, "total_tokens": 6 }}

data holds one entry per input, in order; model is the id you asked for; usage.prompt_tokens is what the request is billed at. The response carries the usual headers, x-token-cost included.

Billing and limits#

An embedding model is paid for input tokens only, per 1M. There is no output or cached bucket.
The request counts against the key's rate limits and spend caps, and holds one concurrency slot while it runs.
It is admitted against the balance exactly as a chat request is, with the same 402 insufficient_balance.
Nothing is cached and nothing is kept: neither the response cache nor any store holds the vector.

Refusals#

guard_credentials reads the input texts before they are sent: an API-key-shaped string is replaced with [REDACTED:<kind>], and the vector encodes the placeholder. guard_injection does not run, because no model reads the text as instructions.

The refusals use the chat codes: 400 invalid_request naming the field (input, encoding_format, dimensions), 400 model_endpoint_mismatch for a chat model, 404 not_found_model when no model is available, 413 request_too_large over 4 MiB, and 429 at a rate limit.