MEGA Hub
New
Docs/Tasks

Embeddings#

The Embeddings endpoint converts one string or a batch of strings into OpenAI-compatible numeric vectors. MEGA routes only to mappings validated for the embeddings task; Embeddings requests do not stream.

Find a compatible model#

Use Inference Models or the authenticated /v1/models endpoint to find a live embeddings route. Compare context length, input price, throughput, Provider, and custom-key support. Output-token price is normally zero for this task.

Call with HTTP#

BASH
curl https://inference.tensorplay.cn/v1/embeddings \  -H "Authorization: Bearer $MEGA_TOKEN" \  -H "Content-Type: application/json" \  --data '{    "model": "BAAI/bge-m3:cheapest",    "input": ["first document", "second document"],    "encoding_format": "float"  }'

The response uses the OpenAI-compatible data array with one indexed embedding per input and a usage object when supplied by the Provider.

Use Python clients#

With the OpenAI SDK:

PYTHON
import osfrom openai import OpenAI client = OpenAI(    base_url="https://inference.tensorplay.cn/v1",    api_key=os.environ["MEGA_TOKEN"],) response = client.embeddings.create(    model="BAAI/bge-m3",    input=["first document", "second document"],)vectors = [item.embedding for item in response.data]

With MEGA InferenceClient:

PYTHON
import osfrom megatensors import InferenceClient client = InferenceClient(provider="auto", api_key=os.environ["MEGA_TOKEN"])vectors = client.feature_extraction(    ["first document", "second document"],    model="BAAI/bge-m3",)

The MEGA client orders returned vectors by their response index.

Preserve vector compatibility#

For production indexes, pin the full model ID, Provider selection, dimensions, and normalization assumptions. Two Providers serving the same Hub model should expose the validated mapping, but Provider-specific revisions or defaults can still affect numeric output.

Store these fields with the index build metadata:

  • Hub model ID and selected Provider policy or slug;
  • vector dimensions and encoding format;
  • application-side normalization or pooling;
  • index creation time and source-data revision.

Do not silently mix vectors from a changed model or dimension in an existing similarity index. Re-embed and rebuild when the vector contract changes.