Location
How Trellis models places, coordinates, and location labels without a location property type.
Three things called "location"
Most systems collapse three different concerns into one column:
- A place: a durable thing with coordinates and a name ("Ferry Building").
- A role: how a record relates to a place ("happened at", "departed from", "going to").
- A label: human text on a record ("Ferry Building, SF"), which may be geocoded on demand.
Trellis keeps them separate. That is the decision in ADR 0063: location is graph-native, not a property type.
There is no location property type
The kernel's property types are a closed set; title, rich_text, number,
select, date, relation, json, and the rest. There is no location,
geo, coordinates, or point.
Adding one would force every editor, renderer, import/export path, and query to grow a geo branch, and it would still not answer the important questions: is this the same place as that one? and what role does it play here?
Places are entities
A place is a record, not an attribute. A Place collection carries:
| Field | Type | Notes |
|---|---|---|
title | title | Display name. |
latitude / longitude | number | WGS84 decimal degrees. |
address | rich_text | Optional human address. |
geometry | json | Optional GeoJSON; validated by the consumer, not the kernel. |
json is deliberate: geometry is opaque to the EAV kernel and must not become a
second schema language. Consumers that need shapes validate them; consumers that
only need pins ignore the field.
A user-defined Place collection is enough today. Promoting a thin core:Place
type later is additive and does not change the model.
Roles are relations
How a record points at a place is a relation:
locatedAt; the general case.origin/destination; travel roles.venue,meetingPoint, and similar roles are user relations.
Cardinality is one place per role by default; multi-place roles use a multi-value relation. That is a schema decision per type, not a kernel change.
Text stays a label
core:Event.location remains rich_text. It is the human-entered label and the
offline fallback. It is not parsed as a coordinate and not used as the
identity of a place.
Geocoding is an adapter and a cache: provider results are keyed by normalized
query, provider, and locale; cached with a TTL; and never written silently into
graph truth. Promoting a geocode result to a durable Place is an explicit
action that creates or links an entity.
Compatibility with flat coordinates
Existing entities with flat latitude / longitude remain valid. The map
extractor treats them as a place-in-progress: if coordinates are present, show
the pin; if not, geocode the label. Linking to a Place entity is additive.
When both exist, the relation wins for identity and the flat fields may be
backfilled or left as a denormalized cache.
What this does not include
Spatial predicates (within, intersects), federated geometry, and
attestation-heavy GIS layers are future query-layer work. They can be built over
json geometry or an index without changing this model.
See the full decision in ADR 0063.