atai-newton-omega-model

Newton Omega Encoder — Time-Series Embeddings via /query

Omega is a time-series encoder: feed it a window of sensor readings, get back a fixed-size embedding you can do ML on. This skill calls the cloud Omega model on the same /query endpoint as the Newton fusion model — one stateless POST per window, embeddings back, no batch job or session lifecycle.

When to Apply

• Embed multivariate sensor windows (vibration, pressure, flow, network, …) into vectors
• Build lightweight downstream ML over those vectors client-side: KNN classification, anomaly scoring, similarity search, PCA/UMAP projection
• Prototype classification without standing up the managed batch pipeline

Do not use this skill when:

• The input is text, an image, or a video — that's the Newton fusion model (/query with Newton::c2_6_8b_fp8_...)
• You need fully-managed, server-side classification over millions of rows

For preparing the raw sensor CSVs (timestamp regularity, gap-aware blocks, temporal-order train/test split, the joint-state feature matrix), see atai-newton-omega-model-data-prep.

The Model

OmegaEncoder::omega_embeddings_1_4

Note the OmegaEncoder:: prefix (not Newton::). Output is a 768-dimensional embedding per channel.

Endpoint

POST {ATAI_API_ENDPOINT}/api/lens/query
Authorization: Bearer <API_KEY>
Content-Type: application/json

Same /query endpoint as the fusion model; the model id selects the Omega encoder. Both ATAI_API_KEY and ATAI_API_ENDPOINT are required — there is no default endpoint, so a wrong-deployment mistake fails loudly at startup. Prod is https://platform.u1.archetypeai.app/api.

Request Shape

Recommended: one request per channel, fanned out in parallel. Each request carries a data.numeric_array event with a single channel — contents is [[/* one channel: w floats */]]. No file_ids, no prompt. Per-channel requests cap the request/response payload size (a single request carrying many channels eventually exceeds REST payload limits and corrupts in transit), and a thread-pool / async fan-out keeps wall-clock close to a single call — see embed() in _common.py.