Translate turn-by-turn directions into GPS coordinates

I’d like to build an assistant that customers interact with to submit orders to my business. We need the customer’s GPS coordinates as part of the order.

I’m in a country where houses aren’t well addressed. So you can’t reliably use, for example, Google’s geolocation API to convert a street address into GPS coordinates.

A standard way to solve this problem would be the following:

  1. Use the assistant’s function call to send the customer an SMS or email containing a URL.
  2. The customer clicks on the URL and is taken to a web page, which asks their permission to share their location.
  3. The page captures their GPS coordinates and sends them back to our application, which we then feed into the assistant.

This would work. But this requires a level of sophistication in digital matters that not all customers have, and forces them to leave our user interface. It also seems so pre-ChatGPT.

What would be really cool is for ChatGPT to be able to follow turn-by-turn directions and estimate the customer’s GPS coordinates. So something like:

  1. The customer says: “From St. Patrick’s church in Coverley, head north on Highway 4. Take the 2nd left. My house is the third house on the right.”
  2. The assistant calls the Google geolocation API to get the GPS coordinates of St. Patrick’s church, which will be a recognisable point of interest.
  3. Given this, ChatGPT “looks” at the map from the starting point, follows the verbal directions, determines the end point, and estimates the end point GPS coordinates, based on knowledge of the map’s scale and distances travelled (and I suppose the earth’s curvature), and maybe using additional calls to Google’s mapping API to “see” different points along the way. I’m speculating.

Obviously, this approach (if it could be made to work), would be heavily dependent on the accuracy of the directions, but it would be no worse than what we do today when we have to physically follow someone’s directions to get somewhere.

I know from experimentation that if you show ChatGPT an image of a map with a labelled starting point and provide it with directions to the end point, it can accurately find the end point on the map. The problem is that without the image, it can’t.

Any thoughts on whether what I’d like to do is remotely feasible? Any comments or suggestions welcome.


If you are building the application then it should be possible to make use of geolocation library calls to get a location, both google and microsoft can usually pinpoint a person to a house even without GPS and if the app is on a mobile device with GPS… you could then include that in your server call from your app.

I think asking the model to predict GPS co-ordinates from a turn by turn is fraught with potential issues.

Regrettably, this sounds right. Thanks.