Agent Guidance, San Francisco Addresses (EAS)

A fixed 5,000-row extract of a 388,550-row live address layer. Two rules prevent most wrong answers. Never extrapolate counts or densities from this sample to San Francisco. Never count rows when you mean addresses, one row is a base address, a unit, and a parcel link, so count(*) is 5,000 while distinct buildings are 4,488.

Query the GeoParquet data asset in place with DuckDB spatial, read_parquet('san-francisco-addresses.parquet'), or load it with GeoPandas. It streams over HTTP range requests, so query the published URL directly rather than downloading first. For rendering use the visual PMTiles asset with its MapLibre styles.

Quirks

• 327 Fulton St appears 18 times, identical address, 18 condo parcel links. Count buildings with count(DISTINCT eas_baseid) and addresses with count(DISTINCT eas_subid).
• Units carry their building's single point, every eas_baseid has exactly one distinct geometry. Deduplicate by base before density maps or clustering.
• parcel_number is null on 1,596 rows, mostly records sourced from the State of CA, so a parcel join silently drops a third of the extract.
• data_updated_at holds the epoch sentinel 1970-01-01 on 2,378 rows, meaning never updated, not 1970. Filter before temporal analysis.
• Nearly every column is text, including latitude, longitude, supervisor, and zip_code. Cast before arithmetic, and read coordinates from geom instead.
• numbertext spells out the supervisor district, not the address number, and supname holds the 2026 incumbents, which dates the snapshot.

CRS

EPSG:4326 degrees, so ST_Distance returns degrees. Use ST_Distance_Sphere for meters, or transform to EPSG:26910 for planar work.

Tested Queries

The grain, records against buildings against units.

Addresses within 250 meters of a point, metric distance done right.

Buildings with the most sampled units.

Related Collections

Nothing joins inside this catalog, the EAS keys point outward. parcel_number joins DataSF's Parcels layer on blklot, cnn joins the street centerlines, nhood joins Analysis Neighborhoods by name. The upstream is a live Socrata endpoint, referenced by URL only. This extract was converted to GeoParquet 2.0 with DuckDB spatial and dates itself, every row carries data_as_of July 24, 2026.