Solution 1

After a few weeks of research I realised my needs were very niche and no API would provide what I needed. I knew Wikipedia would be a great resource but didn't know where to start. After researching the API I found 'On This Day'. Events that stretch back sometimes 2000 years - perfect! Wait, these events don't come with latitude and longitude...

On inspection the events have text with popular figures in pop culture, famous landmarks, and many times country, city or town names. There must be a way to deduce coordinates in this modern age if I have all this information to hand, right? Enter ChatGPT.

The answer was clear. ChatGPT excelled at this. Within these events are famous places, historical landmarks, and well-known figures. With the right approach, there had to be a way to extract locations. ChatGPT could recognize names and infer places even when they weren't explicitly mentioned. It wasn't just matching text - it was reasoning, connecting history, and filling in the gaps where a traditional API would fail. It could adapt, making it the perfect tool to bridge the missing latitude and longitude gap.

Solution 2

ChatGPT costs. After testing, GPT-3 is less expensive at the cost of reasoning and accuracy, struggling with ambiguous locations or missing context. It might recognize a famous landmark but fail entirely to provide its coordinates, even mixing up events (as they are sent in batches). GPT-4, on the other hand, excels with a better understanding of historical events, better disambiguation, and a good grasp of geographic clues. GPT-4 is much slower and costs more with much better results, while GPT-3 is much faster and cheaper with less accurate results. Solution?

I had to use GPT-4 otherwise this was pointless. Streaming the response to my PHP backend was the key to prevent timeouts. How to reduce costs overall? Storing the events was the key, that way a user makes one request for that specific day, I could be almost certain I wouldn't have to send the events to GPT ever again. Enter MySQL. After receiving the events back from ChatGPT I would fill my database up with the events. When a user selects the day again, the events would be ready, and if ChatGPT failed to provide coords for one of the events, I would send a batch of these events again on this new request. This worked surprisingly well!

Solution 3

This one is similar to the second problem and was a more easy solution. Loading the map and searching POIs go towards two seperate API calls that will cost after a certain amount of requests each month. Once again nothing a simple MySQL table can't solve.

I setup a table to track the amount of requests per API and increment the counts after each request. If the app is at the limit the request never makes it to Mapbox and a 429 error is returned to the user letting them know no more requests can be made. To reset the request counts on the 1st of the month I created a scheduled event. Viola, problem solved.