Embeddings

Turn text into vectors for search, retrieval, clustering, and deduplication. The endpoint follows the OpenAI Embeddings API shape, so the OpenAI SDKs' embeddings.create works against it unchanged.

POST https://inference.openrelay.inc/v1/embeddings

The model that serves this family is openrelay/bge-m3, BAAI's BGE-M3: a multilingual embedding model (100+ languages in one vector space) that returns 1024-dimension dense vectors. Check a model's supported_apis on GET /v1/models for embeddings; a model that does not list it answers 400 unsupported_api.

Request

stream is dropped: an embeddings response is always a single JSON body.

Response

{
  "object": "list",
  "data": [
    { "object": "embedding", "index": 0, "embedding": [0.0213, -0.0487, 0.0112] },
    { "object": "embedding", "index": 1, "embedding": [0.0198, -0.0502, 0.0131] }
  ],
  "model": "openrelay/bge-m3",
  "usage": { "prompt_tokens": 18, "total_tokens": 18 }
}

Each embedding has 1024 values; the example is shortened. index is the position of the item in input. usage.prompt_tokens counts every input's tokens, and total_tokens equals it: there is no completion_tokens.

Limits

Split longer documents into chunks of 8192 tokens or fewer, and larger jobs into requests of 1024 items or fewer. A request over either limit is rejected with a 4xx; see Errors.

Billing

openrelay/bge-m3 costs $0.013 per 1M input tokens. It has no output rate, because it generates no tokens. You are billed on usage.prompt_tokens. See Pricing & billing.

Embeddings can also run through the Batch API with endpoint: "/v1/embeddings", at the batch rate.

Examples

curl https://inference.openrelay.inc/v1/embeddings \
  -H "Authorization: Bearer $OPENRELAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openrelay/bge-m3",
    "input": ["How do I cancel my subscription?", "¿Cómo cancelo mi suscripción?"]
  }'