Skip to content

Embeddings

POST /v1/embeddings converts text into vectors for search, retrieval, clustering, and classification. It accepts the same request format as OpenAI's embeddings API, so you can use the OpenAI SDKs by changing the base URL to Keln's. API keys, rate limits, and billing work the same as for chat completions.

from openai import OpenAI

client = OpenAI(base_url="https://api.keln.ai/v1", api_key="sk-keln-…")
resp = client.embeddings.create(
    model="qwen/qwen3-embedding-8b",
    input=["Keln runs open models.", "Paris is the capital of France."],
)
print(len(resp.data[0].embedding))   # 4096
import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://api.keln.ai/v1", apiKey: "sk-keln-…" });
const resp = await client.embeddings.create({
  model: "qwen/qwen3-embedding-8b",
  input: ["Keln runs open models.", "Paris is the capital of France."],
});
console.log(resp.data[0].embedding.length);   // 4096
curl https://api.keln.ai/v1/embeddings \
  -H "Authorization: Bearer $KELN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-embedding-8b",
    "input": ["Keln runs open models.", "Paris is the capital of France."]
  }'

The answer, showing only the first three values of each 4,096-dimensional vector:

{
  "object": "list",
  "model": "qwen/qwen3-embedding-8b",
  "data": [
    {"object": "embedding", "index": 0, "embedding": [-0.0109, -0.0005, -0.0092, …]},
    {"object": "embedding", "index": 1, "embedding": [0.0011, 0.0361, -0.0159, …]}
  ],
  "usage": {"prompt_tokens": 15, "total_tokens": 15, "cost": 0.0000001875}
}

Request

Field Contract
model An embedding model id from GET /v1/models (rows with output_modalities set to ["embeddings"]). Keln returns 400 wrong_endpoint if you use a chat model here or an embedding model with chat completions.
input A string or a list of up to 2,048 strings. Each string produces its own vector, in the same order. Token arrays are not supported (400 token_input_not_supported).
dimensions Optional. Returns a shorter vector than the model's full length, listed on its page at keln.ai/models. Unsupported lengths return 400 embedding_shape_not_supported. Keln returns only the lengths the model produces.
encoding_format float (default) or base64, which encodes each vector as little-endian 32-bit floats. The OpenAI SDKs request base64 and decode it automatically.
user Accepted. Keln removes it before forwarding the request (privacy).

Unsupported fields return 400 invalid_request.

Consistent vectors

Search indexes depend on consistent embeddings, so Keln keeps each model's vectors consistent across requests. If a model's usual source becomes unavailable, Keln continues serving requests, although the resulting vectors may differ very slightly.

Billing

Keln bills embedding models for input tokens only. usage.cost shows the exact charge, based on the model's price in GET /v1/models.

Errors

Code Meaning
token_input_not_supported The input contains token arrays. Send strings instead.
embedding_shape_not_supported The model doesn't support the requested dimensions length.
wrong_endpoint A chat model on this endpoint, or an embedding model on chat completions.
invalid_request An unsupported field or more than 2,048 inputs.
model_not_found, rate and balance errors Same as other Keln endpoints. See Errors & limits.

Keys, budgets and zero data retention work the same as on chat completions.