Data format

A dataset contains items and optional events. IDs are scoped to your workspace.

{
  "name": "store-sample",
  "items": [{"item_id":"sku_42","text":"Merino hiking socks"}],
  "events": [{
    "event_id":"evt_1",
    "person_id":"person_7",
    "item_id":"sku_42",
    "kind":"purchase",
    "timestamp":"2026-09-01T12:00:00Z",
    "value":1
  }]
}

Field	Meaning
item_id	Your stable item key; each event must resolve to an item in the snapshot
text	Item title and description used for text tokenization
event_id	Unique event key within the dataset; duplicates are rejected
person_id	Your pseudonymous account key; identity links must come from your data
kind	The observed action: view, click, purchase, review, or your own name
timestamp	When that action was recorded, RFC 3339 in UTC
value	Optional numeric observation; define its meaning for each event type

The sample import contract accepts up to 1,000 items, 5,000 events, and 1 MiB per request. Larger imports are managed during onboarding. Inputs form an immutable snapshot.

Model calls use item IDs ordered oldest first. The initial next-item head uses sequence order; it does not infer elapsed time or action type from bare IDs. Its model card lists supported inputs.