Pub finder

Find the optimal place where to meet with your friends in Prague.

Try out the Web demo, integrated into Huggingface Spaces 🤗 using Gradio.

How does it work?

We have 1400+ stops in Prague. Given the set of k stops (e.g. Krymská, Anděl, Muzeum) we would like to find a stop (target_stop) that is closest to each of the stops. What is definition of "closest"? Let's define for now the distance function between stop A and stop B as dist(A, B). We can consider case

1. where we minimize the maximum distance from k stops (e.g. worst_case_dist = max(dist(target_stop, Krymská), dist(target_stop, Anděl), dist(target_stop, Muzeum)))
2. or the case where we minimize the total distance from k stops to target_stop (e.g. total_dist = dist(target_stop, Krymská) + dist(target_stop, Anděl) + dist(target_stop, Muzeum)).

Now the simplest solution would be to use dist() function as true distance of the stops considering their GPS coordinates, but we can do better. Since the public transport has different speed of transport depending on the stops themselves I scraped almost all the combinations of the stops in both ways (around 2.1M). I scraped all with arrival date Friday 28.2.2025, 20:00. Now, the distance can be also in minutes that it takes to get from stop A to stop B.

We can now iterate over all stops in Prague and for given list of k stops we can calculate worst_case_dist or total_dist and we can take the best cases. What I am actually doing is that I select top 10 target stops based on geo distance and top 25 target stops based on time distance. Then for all of them I scrape the actual time it takes given the date and time and update the table based on that. I select top 15 target stops and that is the end result.

Usage

Local

uv venv --python=3.12
source .venv/bin/activate
uv pip install -r requirements.txt
python app.py

Now you can visit http://0.0.0.0:3000 and enjoy the app.

Docker

docker build -t pub-finder-app .
docker run -p 3000:3000 --name pub-finder pub-finder-app

Now you can enjoy the app on http://localhost:3000.

To remove the image

docker rm pdf-summarizer

Development

Prepare the geo data of stops scraped from PID.

python3.12 prepare_geo_data.py 

For scraping use the following. Repeat until you scrape all the combinations. Internally you can swith between IDOS and DPP providers. DPP has slightly higher error rate

python3.12 scraping.py --num-processes 50 --num-tasks 50
python3.12 manager.py --threshold-error-rate 0.1
jq '. | (length / 2138906) * 100' results.json

Sources

• https://idos.cz/vlakyautobusymhdvse/spojeni/
• https://pid.cz/zastavky-pid/zastavky-v-praze
• https://mapa.pid.cz/?filter=&zoom=12.0&lon=14.4269&lat=50.0874