Context store
An LDM predicts better when it can see an entity's history, not only its current row. The context store is where a server keeps that history so it is available at inference time.
Building it
The store is built from datasets — you point it at processed datasets and it assembles per-entity context from them.
Its configuration decides how much history each entity carries:
| Setting | What it controls |
|---|---|
| Lookback window | How far back in time to gather events. |
| Max events | The ceiling on events kept per entity. |
| Min context | The minimum an entity needs before it counts as having context. |
| View label | Which dataset key identifies the view. |
| Mismatch strategy | What to do when incoming data does not line up with the stored shape. |
Max events and Lookback window both bound the history, in different units — an entity
busy enough to hit Max events inside the window keeps only the most recent events, not
the whole window.
Operating it
The page reports the store's occupancy and its users, and lists its builds so you can see when context was last assembled and whether it succeeded.
Deleting discards the assembled context. The server keeps serving, but inferences lose the history they were carrying until the store is rebuilt.
Next steps
- Features — what the model reads.
- Batch runs — scoring a population offline.