Whereabouts

A lightweight, fast geocoder for Python using DuckDB. Try it out online at Hugging Face.

Description

Whereabouts is an open-source geocoding library for Python, allowing you to geocode and standardize address data all within your own environment.

Features:

• Two-line installation
• No additional database setup required — uses DuckDB to run all queries
• No need to send data to an external geocoding API
• Fast (geocode 1000s/sec depending on your setup)
• Robust to different errors (typographical, geographic, missing information)

Performance

Whereabouts performs well compared with other geocoders. The charts below show the accuracy when calculated at apartment/unit, house, street, and suburb level, comparing Whereabouts with Google, Mapbox, and Nominatim on sets of residential and retail addresses.

Code to produce these results is found in the whereabouts_testing repo.

Requirements

• Python 3.12+

Installation

Whereabouts can be installed using uv, pip, or conda:

uv add whereabouts

Installation from source

Clone the repo:

git clone https://github.com/ajl2718/whereabouts.git
cd whereabouts

Then create a virtual environment and install dependencies:

uv venv
uv sync

Download a geocoder database or create your own

You will need a geocoding database to match addresses against. You can either download a pre-built database or create your own using a dataset of high-quality reference addresses for a given country, state, or other geographic region.

Option 1: Download a pre-built geocoder database

Pre-built geocoding databases are available from Hugging Face. The list of available databases can be found here.

As an example, to install the small-size geocoder database for Australia:

uv run python -m whereabouts download au_all_sm

Option 2: Create a geocoder database

Rather than using a pre-built database, you can create your own geocoder database if you have your own address file. This file should be a single CSV or Parquet file with the following columns:

Column name	Description	Data type
ADDRESS_DETAIL_PID	Unique identifier for address	int
FULL_ADDRESS	The full address	str
ADDRESS_SITE_NAME	Name of the site. This is usually null	str
LOCALITY_NAME	Name of the suburb or locality	str
POSTCODE	Postcode of address	int
STATE	The state, region or territory for the address	str
LATITUDE	Latitude of geocoded address	float
LONGITUDE	Longitude of geocoded address	float

These fields should be specified in a setup.yml file, which is structured as follows:

data:
    db_name: au_vic_lg
    folder: geodb
    filepath: 'address_file.parquet'
    sep: ","
geocoder:
    matchers: [standard]
    states: [VIC]
schema:
    addr_id: ADDRESS_DETAIL_PID
    full_address: ADDRESS_LABEL
    address_site_name: ADDRESS_SITE_NAME
    locality_name: LOCALITY_NAME
    postcode: POSTCODE
    state: STATE
    latitude: LATITUDE
    longitude: LONGITUDE

addr_id is a unique integer, full_address contains the full address string, while locality_name, postcode, and state are components of the address.

Once the setup.yml is created and a reference dataset is available, the geocoding database can be created:

uv run python -m whereabouts setup_geocoder setup.yml

An example setup.yml file is provided with this repo. Note that the state names listed are specific to Australia and should be changed according to the country's data you are working with.

Geocoding examples

Geocode a list of addresses:

from whereabouts.Matcher import Matcher

matcher = Matcher(db_name='au_all_sm')
matcher.geocode(addresslist, how='standard')

For more accurate geocoding you can use trigram phrases rather than token phrases. Note that you will need one of the large databases to use trigram geocoding:

matcher.geocode(addresslist, how='trigram')

Benchmarking

The matching accuracy of the geocoding algorithms is computed by providing a database, test set, and matching algorithm name. The test set is expected to have columns input_address and best_match.

The benchmarking can be run as:

uv run python -m whereabouts benchmark <database_name> <path_to_test_set> --how <algorithm_name> --threshold <float> --output-csv <path>

For example, for the au_all_sm Australian dataset:

uv run python -m whereabouts benchmark au_all_sm evaluation_data/test_set_au.csv --how standard --threshold 0.5 --output-csv whereabouts/benchmarking_results/benchmarking_results_050726.csv

How the algorithm works

The algorithm employs simple record linkage techniques, making it suitable for implementation in around 10 lines of SQL. It is based on the following papers:

• https://arxiv.org/abs/1708.01402
• https://arxiv.org/abs/1712.09691

Additional changes have been made to the algorithm, including a different similarity function and changes to catch additional errors.

Documentation

Work in progress: https://whereabouts.readthedocs.io/en/latest/

License Disclaimer for Third-Party Data

Note that while the code from this package is licensed under the MIT license, the pre-built databases use data from data providers that may have restrictions for particular use cases:

• The Australian databases are built from the Geocoded National Address File with conditions of use based on the End User Licence Agreement.
• The US databases are still work-in-progress but are based on data from OpenAddresses and so any work with whereabouts based on US address data should adhere to the OpenAddresses license.

Users of this software must comply with the terms and conditions of the respective data licenses, which may impose additional restrictions or requirements. By using this software, you agree to comply with the relevant licenses for any third-party data.

Citing

To cite this repo, please use the following:

@software{whereabouts_2024,
  author = {Alex Lee},
  doi = {[10.5281/zenodo.1234](https://doi.org/10.5281/zenodo.13627073)},
  month = {10},
  title = {{Whereabouts}},
  url = {https://github.com/ajl2718/whereabouts},
  version = {0.3.14},
  year = {2024}
}