Embeddings

Embedding models convert text into a numeric vector that captures its meaning. Use them for semantic search, retrieval-augmented generation (RAG), recommendations, and clustering. The /v1/embeddings endpoint is fully OpenAI-compatible.

Creating embeddings

curl https://gateway.omnia-voice.com/v1/embeddings \
  -H "Authorization: Bearer $OMNIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-Embedding-8B",
    "input": "The quick brown fox jumps over the lazy dog."
  }'

resp = client.embeddings.create(
    model="Qwen/Qwen3-Embedding-8B",
    input="The quick brown fox jumps over the lazy dog.",
)

vector = resp.data[0].embedding

const resp = await client.embeddings.create({
  model: "Qwen/Qwen3-Embedding-8B",
  input: "The quick brown fox jumps over the lazy dog.",
});

const vector = resp.data[0].embedding;

Discover embedding-capable models with GET /v1/models.

Parameters

/v1/embeddings accepts exactly these fields:

Parameter	Description
`model`	The embedding model id (required).
`input`	A string, or an array of strings to embed in one request.
`encoding_format`	`float` (default) for arrays of numbers, or `base64` for a compact encoding.
`dimensions`	Request a reduced output dimensionality (model-dependent).
`user`	An opaque end-user identifier you can attach to the request.

As with chat, unknown parameters are silently ignored; only the fields above are forwarded to the model.

Batch input

input can be a single string or an array of strings. Batching many texts in one request is the recommended way to embed a corpus: it embeds them in a single round trip and reduces requests-per-minute (RPM) pressure against your rate limits.

resp = client.embeddings.create(
    model="Qwen/Qwen3-Embedding-8B",
    input=[
        "First document.",
        "Second document.",
        "Third document.",
    ],
)

for item in resp.data:
    print(item.index, len(item.embedding))

Results come back in data, each with an index that matches the position of the corresponding input, so you can align vectors to their source texts.

The response

{
  "object": "list",
  "data": [
    { "object": "embedding", "index": 0, "embedding": [0.0123, -0.045, "..."] }
  ],
  "model": "Qwen/Qwen3-Embedding-8B",
  "usage": { "prompt_tokens": 11, "total_tokens": 11 }
}

Billing

Embeddings are billed on input tokens only: there are no output tokens, so usage reports prompt_tokens and total_tokens. Batching many inputs in one request is priced the same as sending them individually, while cutting your request count.

Embeddings are the highest-hit-rate surface for response caching: the output is deterministic and RAG pipelines routinely re-embed unchanged text. With caching enabled on your workspace, a repeated identical input is served from cache at 25% of the normal price; the X-Omnia-Cache: hit|miss header shows what happened.

Errors and rate limits

Embedding models have their own per-model TPM/RPM limits; exceeding them returns HTTP 429; see Rate limits. For error codes, see Errors.

Get started

Capture

Observe

Grade & Judge

Evaluate

Improve

Run at scale

Platform

Creating embeddings

Parameters

Batch input

The response

Billing

Errors and rate limits

​Creating embeddings

​Parameters

​Batch input

​The response

​Billing

​Errors and rate limits

Creating embeddings

Parameters

Batch input

The response

Billing

Errors and rate limits