Models & pipelines
The two nouns people mix up — a model is what trains; a pipeline is how and when it trains.
If you remember one distinction from these docs, make it this one:
A model is a trainable artifact. A pipeline is a training workflow.
Models
A model is something that can learn from data: architecture plus configuration. predictAI has four kinds, all rows in the same catalog:
kind | What it is |
|---|---|
custom | Model code authored on the platform (layers, hyperparameters). |
foundation | A pretrained foundation model — predictfm, Chronos-Bolt, TimesFM, and other families. |
byom | A model you trained elsewhere and uploaded. |
A model on its own does nothing. It doesn't know your data, it has no schedule, and it can't serve predictions. It's a blueprint for learning.
Model: "an LSTM with these layers" (what to train)
Pipeline: "train it on daily_sales, nightly" (how & when to train it)Pipelines
A pipeline binds a model to your data and runs trainings:
- a segment — the data view to train on,
- one or more models — a pool; every run trains them all,
- a schedule — cron-like windows, or on-demand training runs,
- a promotion policy — what happens when a training beats the version that's serving.
Pipelines are where automation lives: every model in the pool trains on the same data and competes on score; retrain triggers fire when data drifts; a budget guardrail caps monthly spend; and the promotion policy decides whether winners deploy automatically or wait for your approval.
How they show up in the API
Models live under /v1/models (Models); pipelines under
/v1/pipelines (Pipelines). One model can be used by
many pipelines; a pipeline can reference a foundation model directly without
creating a catalog entry first.
Some route segments contain the word blueprint (e.g.
POST /v1/models) — that's the internal name for model
definitions and appears only in verbatim paths. In prose, they're always
just models.