Glossary · predictAI Docs

Term	Definition
Adapter	A small fine-tuned layer on top of a foundation model, specializing it to your data without retraining the foundation model itself.
Alert	A condition attached to a goal's forecast (threshold, change, anomaly) that fires notifications and realtime events when met.
BYOM	Bring Your Own Model — upload a model trained elsewhere, validate it, and register it in the catalog.
Champion	The best-scoring completed training — the one serving (or eligible to serve) live traffic.
Channel (notifications)	A delivery destination for notifications: email, webhook, and similar.
Channel (realtime)	A WebSocket subscription target: `workspace:{id}`, `goal:{id}`, or `user:{id}`.
Credits	The platform currency. Metered actions (training, inference, discovery) consume credits.
Deployment	A training promoted to live serving; carries status, serving mode, and usage.
Discovery run	One pass of a goal's automation: profile the target, find drivers, run the tournament, deploy the champion.
Driver	A signal the discovery engine found to influence a goal's target, with a lag and strength.
Foundation model	A large pretrained time-series model (any family — predictfm, Chronos-Bolt, TimesFM, …) usable zero-shot or with adapters.
Goal	A declared forecasting outcome (`target_key` + `goal_type`) that the platform automates end to end.
Inference	One prediction request against a deployment (directly, or via a pipeline/goal alias).
Model	A trainable artifact: `custom`, `foundation`, or `byom`.
Model pool	A pipeline's `models` list — every entry trains each run on the same data, and the best score becomes the promotion candidate.
Organization	A group of people and workspaces with shared billing and role-based membership.
Pipeline	A training workflow binding a segment to one or more models with a schedule and promotion policy.
Promotion	The act of putting a training into live serving — creating a deployment.
Promotion policy	A pipeline's rule for handling a winning training: deploy automatically or await manual approval.
Relationship graph	The network of scored, lagged influence edges a goal's discovery builds around its target.
Scenario	A what-if simulation over a goal's relationship graph ("what happens to sales if traffic drops 20%?").
Scope (token)	A permission on an API token, shaped `<resource>.<action>` (e.g. `inferences.create`).
Segment	A curated view over signals — the exact table a model trains on.
Shared / dedicated serving	The two deployment modes: multi-tenant pool vs reserved, explicitly scalable capacity.
Signal	A named time series; the atomic unit of data.
Tournament	A goal's model competition: candidates train on discovered drivers and the best becomes champion.
Training	One run of a pipeline: train, evaluate, score, record.
Workspace	The tenancy container everything belongs to; selected per-request with `X-Workspace-Id`.