Skip to main content
Requires version 2.0.1 or later of the Archetype platform.
The Inference API sends a query to a model that the Archetype Platform hosts and runs for every organization. You don’t build, deploy, or stop anything: name the model in the request body’s model field, and the platform routes the request to a running copy of that model and returns its answer. The API is served under /api/inference and accepts three request formats on the same base URL: Because the OpenAI paths follow the OpenAI wire format, you can point an OpenAI SDK at https://platform.u1.archetypeai.app/api/inference as its base URL and use your Archetype API key.

Choosing an Endpoint

The three endpoints reach the same models: for the same input and settings, they return the same answer. Pick by what you send and what client you already have.
  1. Moving from the Direct Query API, or using Omega: use /inference/query. It takes the body you already send to /api/lens/query, with a new URL and model name. It’s also the only endpoint that applies repetition_penalty, which stops Newton repeating itself in long lists.
  2. Sending images, or already using an OpenAI SDK: use /inference/chat/completions.
  3. Already built on OpenAI’s Responses API: use /inference/responses. Otherwise prefer Chat Completions; it’s the more widely supported format and behaves the same here.
On every endpoint, media go inline in a request body under 2 MiB, images and video frames are resized to 224 × 224 pixels, and a query stops after about 60 seconds. See the current limitations.

Sessions

Requests sent straight to these endpoints share one stream on the model. To give a caller its own values and state on a model, open a session and send the same requests to it under /api/inference/sessions/{session_id}. See Query a Model Session.

Available Models

Each model answers on the endpoints listed for it. Any of a model’s names works in the model field. newton-c-2.6 and newton-fusion-1.0 take text along with video, image, and audio input. omega-embedding is a numeric encoder: it returns embeddings for sensor windows sent as data.numeric_array events.
Not every input works on every endpoint yet. For example, still images and uploaded files are refused on /api/inference/query, and images and video frames are resized to 224 × 224 pixels. See the current limitations for Newton models and for Omega.

Moving from the Direct Query API

If you call Newton or Omega on /api/lens/query today, these guides map your requests to the Inference API and say when to stay on the Direct Query API for now:

Errors

Errors use the standard envelope described in Errors. Every message names the model you asked for.