Embeddings

Last updated September 21, 2026

POST /v1/embeddings — OpenAI's embeddings API, with your own provider keys. Point a codebase indexer (Roo Code, Kilo Code, Continue) or any OpenAI SDK at Tempr and it embeds with OpenAI, Gemini, Mistral, Cohere, Bedrock and more, on the same virtual key, limits and logs as your chat traffic.

Request

POST https://api.temprhq.io/v1/embeddings
Authorization: Bearer tvk_...
Content-Type: application/json

Body parameters

Example

curl https://api.temprhq.io/v1/embeddings \
  -H "Authorization: Bearer $TEMPR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/text-embedding-3-small", "input": ["def add(a, b):", "return a + b"]}'

Response

OpenAI's shape on every provider, one item per input, in input order:

{
  "object": "list",
  "data": [
    { "object": "embedding", "index": 0, "embedding": [0.0023064255, -0.009327292, ...] },
    { "object": "embedding", "index": 1, "embedding": [-0.012154937, 0.0026786416, ...] }
  ],
  "model": "text-embedding-3-small",
  "usage": { "prompt_tokens": 9, "total_tokens": 9 }
}

An OpenAI-compatible provider's answer comes back as the provider sent it, other fields included, unless Tempr has to re-encode its vectors or add the usage it left out. A translated provider's answer (Gemini, Cohere, Bedrock) is built in this shape. Every response carries x-tempr-request-id, and x-tempr-usage-estimated: true when the provider reported no token count: the count in usage, which is also what's billed, is then Tempr's estimate of four characters a token.

Providers & models

The model id picks the provider, as it does for chat. OpenAI-compatible providers get your request as sent, with the model id swapped for theirs; Gemini, Cohere and Bedrock are translated to their own APIs and back. Codebase indexers ask for the model's dimension, so the defaults are listed for the common models.

Anthropic has no embedding models, so an anthropic/… model gets 400 embeddings_unsupported, as does every other provider not listed here (Hugging Face's router, for one, serves chat only). Token-id input to a text-only provider gets 400 unsupported_input. Both come back before any quota is spent; see embeddings errors. Each provider's own limits still apply, such as the most tokens one input may hold.

Input type

Some models embed a search query differently from the documents it searches. Tell them which with input_type, in either Cohere's words or Gemini's:

OpenRouter and custom endpoints get input_type as sent. Other providers don't take it, so it isn't sent to them. Indexers that can't set it embed queries as documents, which works; setting it where you can improves search.

Models list

GET /v1/models?type=embedding lists the embedding models of the providers you have keys for, each marked "type": "embedding" with "architecture": {"output_modalities": ["embeddings"]}; ?type=all lists them beside the chat models, which have "type": "language". Plain GET /v1/models lists chat models only, so a chat model picker filled from it never offers an embedding model. From AWS Bedrock it lists the embedding models you entered with your Bedrock key. Azure deployments and custom models aren't listed as embedding models, because nothing says which of them embed, but you can call them by id.

A bare model id, such as text-embedding-3-small, resolves among those embedding models, the way a bare chat model id resolves among chat models.

What applies

Embeddings go through the same pipeline as chat, with these differences: