API reference

Embeddings

Turn text into vectors for search, clustering and retrieval. POST /v1/embeddings takes OpenAI's request and returns its response, so client.embeddings.create() works with the same key and base URL as chat.

Create embeddings#

POST/v1/embeddings

Request body

modelstringrequired: An embedding model ID from the list below, or GET /v1/models?kind=embedding.
inputstring | arrayrequired: The text to embed: one string, or an array of up to 2,048 non-empty strings to embed in one call. Arrays of token IDs (or of token-ID arrays) are accepted too. Each input must fit the model's context window.
encoding_formatstring: float (the default) returns each vector as an array of numbers. base64 returns it as a base64 string of little-endian 32-bit floats, smaller on the wire. The OpenAI SDKs ask for base64 behind the scenes and decode it, so you get numbers either way unless you ask for it yourself.
dimensionsinteger: A shorter vector of this many numbers, for models that support it. Passed to the model as sent; one that can't do it may ignore it or answer 400 invalid_request_error.
userstring: Accepted and ignored.

The usual headers come back: x-request-id and the rate-limit headers. Rate limits, spend caps, project budgets and model locks apply exactly as they do to chat.

curl https://fullsendai.com/v1/embeddings \  -H "Authorization: Bearer $FULLSEND_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "bge-m3",    "input": ["How do I reset my password?", "Billing runs on the 1st of each month."]  }'

Response

{  "object": "list",  "data": [    { "object": "embedding", "index": 0, "embedding": [-0.0403, 0.0370, -0.0289, …] },    { "object": "embedding", "index": 1, "embedding": [-0.0295, 0.0064, -0.0729, …] }  ],  "model": "bge-m3",  "usage": { "prompt_tokens": 18, "total_tokens": 18 }}

The response#

objectstring: Always list.
data[]array: One {"object": "embedding", "index", "embedding"} per input. index is the input's position in your array; embedding is an array of numbers, or a base64 string with encoding_format: "base64".
modelstring: The FullsendAI model ID you called.
usage.prompt_tokensinteger: Input tokens, as billed. total_tokens is the same number: embeddings have no output.

import base64, structresult = client.embeddings.create(model="bge-m3", input="hello", encoding_format="base64")raw = base64.b64decode(result.data[0].embedding)vector = struct.unpack(f"<{len(raw) // 4}f", raw)  # little-endian float32

Batching and limits#

Up to 2,048 inputs a call. Batching is the cheap way to embed a corpus: one request counts once against your rate limit, whatever its size. Split larger jobs into chunks of 2,048 or fewer.
Each input fits the context window. A longer one is refused by the model with 400 invalid_request_error; chunk long documents before embedding them.
Kinds don't mix. An embedding model called at /v1/chat/completions, or a chat model here, answers 400 invalid_model_kind with the endpoint it belongs to. See Errors.
Failover applies. When a model's host fails, the call moves to its backup at the same price. See Reliability and routing.

Embedding models#

No embedding model is live just now. Ask GET /v1/models?kind=embedding for the current list.

GET /v1/models lists chat models by default, since tools fill their model pickers from it. Add ?kind=embedding for these, or ?kind=all for everything; see Models.

curl

curl "https://fullsendai.com/v1/models?kind=embedding" -H "Authorization: Bearer $FULLSEND_API_KEY"

Pricing#

You pay for input tokens only, at the model's input price per million; there are no output tokens. The tokens billed are usage.prompt_tokens. A call costs at least one micro-dollar, and each one shows on its row in Logs with its cost. Plan allowance is spent first, then your prepaid balance, as with chat; see Pricing and billing.

Stuck? Email [email protected]System status Changelog