Paano ginawa?

How it works

Saan Tayo? tags every city and municipality in the Philippines by what you can do there, so you can find places you've never heard of. Here's how the tags are made, how accurate they are, and what it cost.

1. What's on the map

All 1,642 cities and municipalities, from the Philippine Statistics Authority's official list. Each one is tagged for 13 vibes: beach, island hopping, diving, surfing, mountains, waterfalls, caves, hot springs, cool climate, countryside, heritage, festivals and food.

1,092 towns have at least one vibe. The rest either have none that the sources describe, or too little written about them to say.

2. How a town gets its tags

1. The town list. The PSA's official list. Every source is matched to a town by its official code (PSGC), never by name: 329 towns share their name with another town (there are 9 San Joses).
2. What's written about it. The town's Wikipedia article and its Wikivoyage travel guide, the Wikipedia articles of places in the town (islands, falls, churches, parks), and a Wikivoyage guide to its island or small province if there is one. All trimmed to the parts about places, nature, food and culture.
3. The questions. An AI model, Jev, reads that text and answers one yes-or-no question per vibe, giving how likely "yes" is.
4. The rules. Code, not the model, turns those answers into the tags you see.
5. Map facts. OpenStreetMap counts of beaches, peaks, waterfalls and more are shown as context. They never set a tag.
6. Photos. From Wikimedia Commons, credited to each photographer. Towns without a good photo get an illustration, labelled as one.

3. The model judges, the code decides

Jev, a model by TypeSafe, answers questions like these for each town, using only the text it was given:

Does the town have a beach that visitors go to?

Does the town hold a festival that visitors come to?

Does the town have historic structures or heritage-listed sites that visitors can visit?

Everything after that is plain code, so it's the same for every town:

• Yes when the model's confidence is 75% or more, maybe from 50%. The words on the site follow it: Strong from 90%, Good from 75%, Maybe from 50%.
• Not enough info when the text about the town itself is very short (under 250 tokens, about 190 words). Then only a clear "yes" is shown, never a "no". A guide to a whole island doesn't count toward this, because it's about several towns.
• No sea vibes inland. Beach, island hopping, diving and surfing are shown only for coastal towns.
• The main vibe on a card is simply the highest one.
• For coastal towns, the sea vibes also read the titles and descriptions of the town's photos on Wikimedia Commons, because many beaches are photographed but never written about.
• Two small hand-kept lists guard against crediting a town with its neighbour's places. One says which towns each island or area guide is read for (only guides covering 10 towns or fewer). The other removes sentences that clearly describe a different town, such as a line about Samal Island in Davao City's article. Neither list ever sets a tag.

4. How accurate it is

We checked the tags against a reference set of 150 towns. The reference labels were made with the help of a separate AI assistant using web search, plus the author's own knowledge of some towns, without looking at the model's answers. So these figures are agreement with that reference set, not verified ground truth.

The bar for a vibe is 85% agreement when it says yes. A missed town only hides it from one filter, but a wrong tag sends someone somewhere for the wrong reason. Some things to know:

• * Cool climate is under the bar: 80.0% (4 of 5). Only 5 reference towns show it as yes, and one of them disagrees. It is shown anyway, so treat it with care.
• Hot springs is at 86.7% (13 of 15), just over the bar. Its reference "no" labels are weaker than the others: for half the towns "no" means a web search found no hot spring, not that one was ruled out.
• Mountains is at 87.5%. Its mistakes are mostly towns whose text mentions only foothills or "mountain people".
• Countryside is the weakest vibe. It catches 4 of 15 reference towns, and with so few examples its agreement figure rests on 4 towns.
• Festivals and Food are rarely wrong but catch fewer than half the towns. Articles seldom name a festival visitors travel for, or a specific dish.
• Most vibes catch about half the towns that have them. The tags show what is written down, and much isn't.
• 21% of towns (347) are marked "not enough info", because little has been written about them.

5. What was tried and changed

• Dropped "City life". One mall or a few restaurants was enough for a "yes", so only 1 in 3 of its yes tags agreed with the reference.
• Widened "Heritage" to include heritage-listed sites, not just old buildings.
• Asked two ways. Each question was tested with and without a definition. Beach and mountains work better without one; the rest work better with it.
• Didn't trust the model's own "enough info". It flagged half the country as too thin. A simple text-length rule matched the reference far better.
• Tested map-based tags and dropped them. No rule, such as "the map shows a waterfall", reached the 85% bar. So the map is context only.
• Fixed missing beaches. 98% of coastal towns with a missed beach had no beach anywhere in their text. Adding their Commons photo titles raised coastal towns tagged Beach from 253 to 348, with agreement staying above the bar.
• Read more about each town. A famous place is often described in its own article, not the town's: Boracay is in Malay, and Malay's article barely mentions diving. Reading the articles of places in each town, guides to small islands and provinces, and more of the longest city articles raised the matches on a list of 212 things well-known destinations are known for from 181 to 190. It changed 444 towns' text, and every vibe had to stay above the bar on the reference set to be kept.
• Kept Countryside as it was. With the longer text, one rice field beside a highway was enough for a "yes", and its agreement fell to 3 in 4. So Countryside still uses its earlier answers.
• Grew the reference set from 70 to 150 towns, so that surfing, caves, hot springs, cool climate and countryside could be scored at all.
• Checked other languages. Tagalog, Cebuano and other Wikipedias have an article for every town, but most are short automatic stubs, so they weren't used.

6. What it cost

• About US$0.51 in AI costs for all 1,642 towns: $0.26 for the main tagging and photo choice, $0.07 for the beach fix, and $0.18 for reading more text about 444 towns.
• About 3,200 tokens per town (a token is about ¾ of a word), read once per town. The model counts the text once for all 15 questions.
• Nothing runs when you visit. The tags are made ahead of time; the site makes no AI calls.
• Everything else is free: Wikipedia, Wikivoyage, Wikidata, Wikimedia Commons, OpenStreetMap, the PSA list, and hosting on Vercel's free plan.

7. Limits

• Tags reflect what's written, not what's there. A town with a great unwritten-about waterfall won't show it.
• A shared guide can blur towns. A guide to a small island group is read for each town it names. A general line in it ("fine beaches and dive sites") can count for every one of them.
• 8 towns appear as dots, because no border shape exists for them yet. A few small gaps between borders are shown as unassigned land.
• Map facts come from OpenStreetMap, which volunteers keep. Counts can be off, and a feature mapped twice may be counted twice.
• Distances are straight lines ("as the crow flies"), not travel times.
• Photos are matched to towns by their Commons category and checked by the model, so an occasional photo may show a nearby place.

8. Credits

Every source, licence, article revision and photographer is on the Credits page. The data keeps its sources' licences.