Pipelines

Training workflows — bind a model to your data, on a schedule, with automation.

A pipeline is a training workflow. It binds together everything a training run needs — what to train, on what data, when, and what to do with the result:

Pipeline = segment  (the data view)
         + model(s) (a pool of one or more models —
                     every run trains them all, best score wins)
         + schedule (cron-like windows, or run-once)
         + policy   (auto-promote the winner, or ask you first)

You need a pipeline whenever you train on the platform. Every training run belongs to a pipeline; every deployment traces back to one.

Thirty seconds of API

# Create: race a foundation model against your own model on a segment.
# The models array takes any mix of catalog IDs — every run trains
# them all; best score wins.
curl -X POST "$API_BASE/v1/pipelines" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Daily sales forecaster",
    "segment": "SEGMENT_ID",
    "workspace_id": "'$WORKSPACE_ID'",
    "models": [
      { "family": "predictfm", "slug": "predictfm-f-v0.1" },
      { "model": "MY_CUSTOM_MODEL_ID" }
    ]
  }'

# Trigger a run right now
curl -X POST "$API_BASE/v1/pipelines/$PIPELINE_ID/train" \
  -H "Authorization: Bearer $TOKEN" \
  -H "X-Workspace-Id: $WORKSPACE_ID"

In this section

Concepts

Anatomy, schedules, the model pool, and promotion policy.

Create & manage

Create, list, inspect, edit, and delete pipelines.

Train now

Trigger a run on demand — and what can stop one.

Promotion decisions

Approve or reject a winning training when policy is manual.

Trainings themselves (run history, reports, costs) live in Trainings; putting a winner live is Deployments.