Embeddings

Embeddings convert text into vectors whose distance reflects similarity. They are the basis of semantic search and retrieval-augmented generation.

Use an embedding model — one carrying the embedding category in the catalog. Chat models cannot serve this endpoint.

cURL

curl https://api.resetdata.ai/api/v1/embeddings \
  -H "Authorization: Bearer $RESETDATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "EMBEDDING_MODEL",
    "input": "Sovereign AI infrastructure in Australia."
  }'

Python

resp = client.embeddings.create(
    model="EMBEDDING_MODEL",
    input="Sovereign AI infrastructure in Australia.",
)
vector = resp.data[0].embedding

JavaScript

const resp = await client.embeddings.create({
  model: "EMBEDDING_MODEL",
  input: "Sovereign AI infrastructure in Australia.",
});
const vector = resp.data[0].embedding;

Replace EMBEDDING_MODEL with an embedding model slug from the catalog.

Batching

input accepts an array, and batching is markedly more efficient than one request per string — fewer round trips, and one admission through the queue instead of many.

resp = client.embeddings.create(
    model="EMBEDDING_MODEL",
    input=["First document.", "Second document.", "Third document."],
)
vectors = [d.embedding for d in resp.data]

Results come back in data, each with an index matching your input order. Match on index rather than assuming ordering.

Things to get right

• Use the same model for both sides. Vectors from different models are not comparable. If you re-embed your corpus with a new model, you must re-embed your queries too.
• Chunk before embedding. Embedding models have short context windows relative to chat models, and a vector over a whole document is too diffuse to retrieve well. Paragraph-to-page sized chunks with slight overlap is a reasonable default.
• Check for asymmetric models. Some embedding models expect queries and documents to be embedded differently, via a parameter such as input_type. Check supported_parameters — see Model parameters.
• Store the model and dimension alongside your vectors, so a future migration is a known quantity.

Cost

Embeddings are charged on input tokens only — there is no generated output — and are typically far cheaper per token than chat models. Embedding a large corpus is still a bulk operation worth estimating before you start; see Usage and costs.

Improving retrieval

Embedding similarity alone is often not precise enough at the top of the ranking. The usual fix is to retrieve generously with embeddings, then reorder with a reranker before passing results to a model.