Embeddings

Embeddings turn text into vectors whose distance reflects meaning: the basis of semantic search, retrieval-augmented generation, clustering and deduplication.

API reference: Embeddings

The model

arklabsai/bge-m3 produces 1,024-dimensional vectors for more than 100 languages, Polish and the other EU languages included, from inputs of up to 8,192 tokens. Vectors are L2-normalised, so cosine similarity is a dot product.

Create embeddings

response = client.embeddings.create(
    model="arklabsai/bge-m3",
    input=[
        "Umowa najmu lokalu użytkowego",
        "Commercial lease agreement",
        "Recipe for pierogi ruskie",
    ],
)
vectors = [item.embedding for item in response.data]
print(len(vectors), len(vectors[0]))   # 3 1024

const response = await client.embeddings.create({
  model: "arklabsai/bge-m3",
  input: ["Umowa najmu lokalu użytkowego", "Commercial lease agreement", "Recipe for pierogi ruskie"],
});
const vectors = response.data.map((d) => d.embedding);

curl https://api.ark-labs.cloud/api/v1/embeddings \
  -H "Authorization: Bearer $ARK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "arklabsai/bge-m3",
    "input": ["Umowa najmu lokalu użytkowego", "Commercial lease agreement"]
  }'

{
  "object": "list",
  "data": [
    { "object": "embedding", "index": 0, "embedding": [0.0123, -0.0456, …] },
    { "object": "embedding", "index": 1, "embedding": [0.0119, -0.0398, …] }
  ],
  "model": "arklabsai/bge-m3",
  "usage": { "prompt_tokens": 22, "total_tokens": 22, "credits_used": 1, "credit_balance": 49987 }
}

Parameters

Example: semantic search

import numpy as np

docs = ["Faktura VAT musi zawierać NIP sprzedawcy.", "The office is closed on public holidays.", "Invoices must show the seller's tax number."]
doc_vecs = np.array([d.embedding for d in client.embeddings.create(model="arklabsai/bge-m3", input=docs).data])

query = client.embeddings.create(model="arklabsai/bge-m3", input="What must an invoice contain?").data[0].embedding
scores = doc_vecs @ np.array(query)            # vectors are normalised: dot product = cosine
for i in scores.argsort()[::-1]:
    print(round(float(scores[i]), 3), docs[i])

The Polish and English sentences about invoices score close to each other: bge-m3 maps languages into one space, which makes cross-lingual search work without translation.

Limits and usage

• An input longer than 8,192 tokens returns 400 "This model's maximum context length is 8192 tokens." Split long documents into chunks.
• A batch is processed as one block: prompt_tokens is the number of inputs times the length of the longest input. Group inputs of similar length to use your credits well. A batch that is too large returns 400 "Batch too large"; send fewer or shorter inputs per request.