Segment-based inference

/infer-segment is the production path: instead of building an input tensor yourself, the platform rebuilds the exact input the model saw during training, from the segment's live data, on every request.

curl -X POST "$API_BASE/v1/inference/deployments/$DEPLOYMENT_ID/infer-segment" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID" \
  -H "Content-Type: application/json" \
  -d '{}'

from predictai import PredictAI

client = PredictAI(token="pa_live_…", workspace_id="ws_…")

data = client.inference.post_inference_deployments_by_deployment_id_infer_segment(DEPLOYMENT_ID, json={})

import { PredictAI } from "@predictai/sdk";

const client = new PredictAI({ token: "pa_live_…", workspaceId: "ws_…" });

const data = await client.inference.postInferenceDeploymentsByDeploymentIdInferSegment(DEPLOYMENT_ID, { json: {} });

The same route exists on the pipeline and goal aliases.

What the platform prepares

For a deployment backed by a trained model, each request:

1. Looks up the model's configuration — segment, sequence length, feature set.
2. Fetches the segment's most recent live data.
3. Applies the normalization and engineered features used during training.
4. Assembles the input in the exact shape the model was trained on.
5. Runs inference, inverse-scales the outputs back to real units, and maps them to segment labels.

That's why an empty body {} is a complete request — everything the model needs is derived from the deployment. The response's data_period tells you the window of data it used, and feature_names the features that went in.

Body options

For trained-model deployments the body is optional. One debugging knob is supported:

Foundation deployments (a foundation model deployed zero-shot, with no trained model attached) have no bound segment, so you tell them what to forecast:

curl -X POST "$API_BASE/v1/inference/deployments/$DEPLOYMENT_ID/infer-segment" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID" \
  -H "Content-Type: application/json" \
  -d '{
    "segment_id": "SEGMENT_ID",
    "horizon": 28,
    "quantile_levels": [0.1, 0.5, 0.9]
  }'

from predictai import PredictAI

client = PredictAI(token="pa_live_…", workspace_id="ws_…")

data = client.inference.post_inference_deployments_by_deployment_id_infer_segment(DEPLOYMENT_ID, json={
    "segment_id": "SEGMENT_ID",
    "horizon": 28,
    "quantile_levels": [0.1, 0.5, 0.9],
})

import { PredictAI } from "@predictai/sdk";

const client = new PredictAI({ token: "pa_live_…", workspaceId: "ws_…" });

const data = await client.inference.postInferenceDeploymentsByDeploymentIdInferSegment(DEPLOYMENT_ID, {
  json: {
    segment_id: "SEGMENT_ID",
    horizon: 28,
    quantile_levels: [0.1, 0.5, 0.9],
  },
});

Omitting segment_id on a foundation deployment returns 400 VALIDATION_ERROR.

Reading the response

The full shape is shown in Run a forecast. The segment-specific fields:

When to prefer raw /infer

Reach for raw inference when segment preparation gets in the way:

• You compute features in your own system and want the model to see exactly those values — not the segment's live data.
• You're backtesting or replaying historical inputs.
• You're integrating an existing client that already speaks a standard {"input": …} tensor format.

You take on preprocessing yourself: raw inputs must match the training feature order, and outputs come back unscaled under the synthetic output label. See Run a forecast.