pvv5385/overture-usa-places-embeddings

Overture Places Embeddings — USA Multi-City (pilot)

What is this dataset?

1,619,134 points of interest (restaurants, shops, services, parks, etc.) from 13 US cities, sourced from the Overture Maps Foundation Places theme. Each record has been turned into a short natural-language description ("canonical text") and embedded with BAAI/bge-base-en-v1.5, so the dataset can be dropped directly into a semantic search, RAG, or recommendation pipeline without any additional text processing or embedding step.

This is a pilot for a planned top-100-US-cities-by-population dataset. It currently covers the 10 largest US cities by population plus San Francisco, Oakland, and Berkeley (carried over from the earlier SF Bay Area v1 dataset — see pvv5385/overture-sf-bay-area-places-embeddings, which remains published separately and unchanged). More cities will be added incrementally; see docs/MULTI_CITY.md in the GitHub repository for the extension process.

Cities included (pilot)

City	State	Records
New York	NY	429,424
Los Angeles	CA	314,748
San Diego	CA	89,288
Houston	TX	198,543
Dallas	TX	105,102
San Antonio	TX	72,028
Chicago	IL	103,352
Phoenix	AZ	133,603
Philadelphia	PA	62,880
San Jose	CA	46,647
San Francisco	CA	43,420
Oakland	CA	14,327
Berkeley	CA	5,772

Why was it created?

To scale the original SF Bay Area starter dataset into a much larger, ongoing multi-city semantic-search dataset, organized so more cities, states, and countries can be added over time without restructuring.

Source

• Upstream data: Overture Maps Foundation, Places theme, release 2026-08-19.0, fetched directly from the public overturemaps-us-west-2 S3 bucket.
• Geographic scope: per-city bounding boxes; see data/registry/us_cities.csv for exact coordinates.

License

Overture's Places theme is published under CDLA-Permissive-2.0 and Apache-2.0, and contains no OpenStreetMap-derived data, so it carries none of ODbL's share-alike obligations. This dataset (canonical text + embeddings, a derivative of Overture Places data) is distributed under the same terms: CDLA-Permissive-2.0. You may use, modify, and redistribute it, including for commercial purposes, without an obligation to share derivative works under the same license. Attribution to the Overture Maps Foundation is appreciated but not required by the license; see Overture's attribution guidance for details.

This dataset was not derived from the Yelp Academic Dataset or any other Yelp data, and no Yelp terms of use apply to it.

How was it cleaned?

Raw Overture place records were processed with scripts/build_region.py (which wraps the same cleaning logic as v1's scripts/prepare_overture.py), which:

1. Filters out records below a confidence threshold of 0.7.
2. Drops records missing required fields (name, category, coordinates).
3. Flags (but keeps) low-information records — those with very little descriptive metadata beyond a name and category — via the low_information column, so downstream users can choose to exclude them.
4. Renders each surviving record into a canonical natural-language sentence (name, category, neighborhood/locality, and up to 2 alternate categories) per the template in docs/CANONICAL_TEXT_OVERTURE.md.

How were embeddings generated?

Canonical text was embedded with BAAI/bge-base-en-v1.5 via sentence-transformers, with normalize_embeddings=True (so cosine similarity is a plain dot product). See scripts/embed.py.

Schema

Column	Type	Description
id	string	Overture Maps GERS place ID
canonical_text	string	Natural-language description that was embedded
embedding	list<float32>[768]	L2-normalized bge-base-en-v1.5 embedding
category_primary	string	Overture primary category (e.g. coffee_shop)
category_alternates	list<string>	Up to 2 alternate categories
brand	string | null	Brand name, if the place is a chain
locality	string	City/neighborhood
region	string	State code
country	string	Country code (US)
lat / lon	float	Coordinates
confidence	float	Overture's confidence score for the record
website / phone / socials	string | null	Contact metadata, where available
low_information	bool	True if the record had minimal descriptive metadata
source_dataset	string	overture
source_version	string	Overture release version used

Files in this repo

• overture_usa_embedded.parquet — combined file, all cities.
• usa/<state_code>/<city-slug>/embedded.parquet — per-city files, for users who only want a subset of cities.
• registry/us_cities.csv — the city registry (bounding boxes, population, rank) used to build this dataset.
• benchmark/queries.json — retrieval benchmark queries and rule-based ground truth (category-membership), reused from v1.

How can it be used?

import pyarrow.parquet as pq
import numpy as np
from sentence_transformers import SentenceTransformer

table = pq.read_table("overture_usa_embedded.parquet")
embeddings = np.array(table.column("embedding").to_pylist(), dtype=np.float32)

model = SentenceTransformer("BAAI/bge-base-en-v1.5")
query = model.encode("a quiet place to get coffee and read", normalize_embeddings=True)

scores = embeddings @ query
top5 = np.argsort(-scores)[:5]
for i in top5:
    print(table.column("canonical_text")[int(i)].as_py())

Filter by locality/region/country, or load a single city's usa/<state>/<slug>/embedded.parquet file, to scope to one metro area.

See the GitHub repository for the full pipeline (fetch, clean, embed, benchmark) and docs/MULTI_CITY.md for how to add more cities.

Limitations

• Pilot scope: 13 cities, not yet the full top-100 by population — this is an in-progress dataset that will grow over time.
• Snapshot in time: reflects the Overture 2026-08-19.0 release; businesses close, move, and rename, and this dataset is not kept in sync with live data.
• Category coverage skew: category distribution follows whatever OSM/Overture source data covers well in each region; some categories (e.g. niche services) may be sparse or absent, and coverage quality may vary city to city.
• low_information records: included but flagged, not filtered — they carry weaker text and may embed less usefully; filter on this column if you need cleaner precision.
• Single embedding model: only bge-base-en-v1.5 embeddings are published. If you need a different embedding model, re-embed canonical_text yourself (see scripts/embed.py).