มดแดง Mot Dang — Chiang Mai · Chiang Rai city directory

Thai-first, 1997-directory-genre city index + GIS. The dataset is the product; the directory pages and (later) the map are two lenses on it. Chiang Mai is the full build-out; Chiang Rai is a wireframe that grows.

Quickstart

pip3 install --user qrcode        # once — build.py inlines a QR per place page
python3 importers/import_all.py   # fold source corpora -> data/canonical/{cm,cr}.json
python3 build.py                  # -> docs/ (GitHub Pages ready, .nojekyll included)
open docs/index.html              # works from file://, offline

Or use the Desktop launcher: Mot Dang.command. Without qrcode installed, build.py still runs fine — QR boxes are just skipped.

Ground rules

• Thai canonical, EN a display layer (client-side toggle; more languages later).
• Category tree is data — data/categories.json. Empty categories are hidden by the build.
• A wireframe shelf means "no data" — check it does not mean "wrong rule". A child whose match names a sub nobody emits renders identically to one genuinely waiting for data — muted, 🐜 มดกำลังไปเก็บ — so a typo reads to every visitor as "Chiang Mai has no tattoo studios". It hid 35 studios, 56 salons and 40 vegetarian kitchens that were in the data the whole time. tests/test_facets.py now fails on a rule that matches nothing and on records that reach no shelf; shelves deliberately awaiting data are listed in KNOWN_EMPTY, so leaving one empty is a decision someone wrote down rather than an accident.
• A KNOWN_EMPTY reason can be wrong, and nothing checks it. beauty/salon sat on that list with the reason "no OSM signal separates a salon from a hairdresser" — true about the tags, false about the shops. เสริมสวย is THE Thai word for a women's salon and it was in sixty-two names the whole time; the barber shelf next to it showed 6 of the city's 62 for the same reason. Both were fixed by reading the names, in Thai, in one afternoon (WO-22, importers/audit_beauty.py). A written-down decision still put two lying shelves in front of readers for months, because the test can only ask "is this shelf empty on purpose", not "is the purpose still true". Before a shelf is declared unfillable, read the names — in the language the shop wrote them. The corollary holds too: when the names genuinely say nothing, say so with a number. Across all 18,686 records not one shopfront names a perm, an updo, textured hair or a house call, and audit_beauty.py prints those four zeros every run so the hole stays visible instead of being mistaken for a rule nobody got round to writing.
• A shelf's name is a promise. repair/home was called ช่างบ้าน-ประปา-ไฟ and held no plumbers and no electricians, because OSM maps none here — so it was renamed to what it actually holds. Likewise community/intl-clubs read ชมรมนานาชาติ over a Thai school's alumni association and a village hall, and is now ชมรม-สมาคม. If the data cannot keep the label's promise, change the label, not the reader's expectations.
• Field/curated truth beats crawled truth — records in data/curated/ always win over a crawl refresh (same discipline as mueang-map).
• Provenance on every page — source type + fetch date shown to readers.
• Type, page data and the basemap are self-hosted and baked — vector tiles AND label glyphs come from our own bucket, configured in map_shell.py. The four Noto Sans PBF ranges went into assets/glyphs/ on 2026-08-20.
• No local paths in published output — grep -rl "/Users/" docs/ is checked before every deploy. Publish as NaNoBotCo.

Data sources (all local, zero network)

source	records	lands in
thai-answers/catalog.db	316 massage venues	cm/นวด-สปา
cm-womens-health osm-health.json	258 facilities	cm/หมอ + ของจำเป็น
mueang-map canonical osm.json	378 lens points	cm/วัด ร้านอาหาร ที่เที่ยว
mueang-map osm-chiang-rai.json	289 wats	cr/วัด (wireframe seed)
data/curated/featured-chiang-rai.json	hand-entered field truth	cr featured
mueang-map's Commons metadata (media field)	114 real, licensed wat photos	assets/photos/ — see importers/import_photos_commons.py
cache/overpass/*/fixtures.json	833 ATM/toilet points	facets on 321 shops — see importers/import_fixtures.py
cache/overpass/*/{crafts,community,business}.json	222 elements	laundries, coworking, clubs, trades — the three categories that had shelves but no query

Facets — telling one branch of a chain from the next

A crawl gives 386 7-Elevens identical records: same brand, same wikidata id, same website, 386 pins. But nobody chooses a 7-Eleven by its brand. They choose the one with the cash machine, or the bake-off oven, or somewhere to sit — and none of that is in OpenStreetMap. Of 385 7-Elevens in the snapshot, two carried has:slurpee, one carried amenity=atm, one carried internet_access. Mapping them better is a data-collection problem wearing a crawl's clothes.

• data/facets.json is the schema — 13 facets, Thai/EN labels, an icon, and the question to ask a passer-by. Same discipline as the category tree: the list is data, not code. appliesTo matches record.sub, so nothing here is 7-Eleven-specific — pharmacies and fuel stations inherit the row when their turn comes.
• Presence is the only claim. An absent facet renders as nothing rather than as "no". We know 109 shops have an ATM; we do not know the other 406 lack one, and a directory that implied it would be lying quietly.
• The border is the provenance. Dashed = joined by distance from a map point. Solid = tagged in OSM. Solid and bold = a person stood there. A reader can tell how much to trust a tag before they open the tooltip.
• importers/import_fixtures.py works two seams the shop record itself does not have. OSM maps an ATM as its own node beside the store: joining amenity=atm within 30m takes the known ATM count from 1 to 109. The hit curve flattens past 30m (73 at 20m, 79 at 30m, 87 at 50m) — the extra hits at 50m are bank lobbies across the road, so 30m is where evidence stops and guessing starts.
• Filter chips are AND, not OR — the question is always "a cash machine and somewhere to sit". The heading count follows the filter, so it says what is on screen.
• The vocabulary lives in three files (schema, worker, build) because a Cloudflare Worker cannot read the repo at request time. tests/test_facets.py fails if they drift: an unrecognised key is silently discarded by the worker, so a contributor would tick "has a bakery", get a success message, and lose the fact.
• Ticks do not open a claim. Claiming locks a record against everyone else, so letting a passer-by claim by ticking a box would let a stranger lock a shop out of its own listing. A claim still needs a real way to reach the shop; ticks only ride along.

🏷 Tags — what a place ALSO is, across every shelf

A shelf says what kind of place this is (one branch of the tree); a facet says whether that branch is worth walking to; a tag says what the place also is, across every shelf — vegan, bitcoin, wifi, wheelchair, open 24 h, inside the moat, a 7-Eleven, a royal temple. Each tag is a page a reader lands on for "vegan chiang mai": the exhaustive-list queries the tree cannot answer.

• data/tags.json is the schema — ~70 tags in 13 families, Thai/EN, a glyph, and exactly ONE rule over a field the record already carries (an attrs value, a facet key, the honours list, the moat polygon). No tag without a rule; no rule without a source field; nothing inferred from a name. The vocabulary was mined from the catalogue first (140 attrs keys over 16k records) — a tag exists only where the data already answers for it. Brand tags are generated per chain from attrs.brand through the same _brand_index the brand shelves use.
• Provenance travels. tags_layer.py assigns once after load(); each record in data/places.json carries tags + tagVia (attr: · facet:: · moat · honours: · brand · curated:), and the pill's tooltip on the place page says the same in words.
• Empty is hidden by design. A tag page (/<prov>/tag/<slug>.html) needs min_tag records in that province; a tag×shelf page (<slug>--<shelf>.html) needs tag_shelf_min. Counts are per province, unmerged. Index at /tags.html, Yahoo-style Tag (count) by family; the search index matches both names of every tag a place earned.
• Curated outranks derived. data/curated/tags_curated.json — {"tags": {"<slug>": {"th","en","glyph","family","ids":[…]}}} — for a person's own list; applied first, provenance curated:<slug>.
• tests/test_tags.py fails if a defined rule matches nothing anywhere (unless named in KNOWN_EMPTY with a reason), if a page's count disagrees with places.json, if a page exists below threshold, or if a tagged place page has no pills.

The 1997 layer

• Search + services bar on every page; subcategory shelves with counts; empty shelves show muted with 🐜 มดกำลังไปเก็บ.
• my.html — the personal start page: pin shelves, get "+N new" badges since your last visit, a daily pick drawn from your pins, custom links, sticky notes. All localStorage; Mot Dang follows no one around.
• Sorting: ก→ฮ or 📍 ใกล้ฉัน (client-side geolocation, nothing leaves the device).
• Sharing: pill-button row — native Web Share (mobile), LINE, WhatsApp, Telegram, copy-link — on every page type; place pages also get an inline QR code (base64 PNG, zero extra requests) to scan or print by a door.
• Photos: an original hand-drawn wat illustration is the default image everywhere a real photo is missing (deliberately — see build.py's WAT_SVG). assets/photos/<record-id>.jpg overrides it automatically; 114 real, Wikimedia-Commons-licensed wat photos already populate this from mueang-map's existing crawl (proper credit line rendered from assets/photos/credits.json).
• Sponsors: rotating 1997-innocent ad boxes from data/ads.json, always marked ผู้สนับสนุน; policy on advertise.html.
• Contact drive: pages without phone/LINE/FB carry a "tell the ants" CTA that pre-fills a GitHub issue with the place id. Contact info is the directory's real currency — capture it everywhere.

Route planner (plan.html)

The errand-run layer, for the person this site is actually for: someone on foot or on a scooter who wants a noodle stand near a clinic near a nail place, this afternoon.

• Add a stop from anywhere — a ring button on every listing row, a labelled pill on every place page. Both drive one md-plan list in localStorage, so the nav chip's count is live on every page. Cap is 8 stops.
• The plan lives in its URL — plan.html?stops=cm:<slug>,cm:<slug>,… in visiting order. That link is the plan — no server state — so it can be handed to somebody over LINE. Arriving by a shared link does not overwrite the reader's own plan silently — a banner offers to keep it.
• One map, drawn in Python-free JS — same inline-SVG approach as the events map, no library: numbered pins, both routed lines, a scale bar, a legend, and the moat for orientation. (Not yet wired to the basemap shell; when it is, the drawing stays exactly as it is and gains streets underneath.)
• Routed on real streets, per mode. See below — this is the part worth knowing about.
• Out: reorder by hand or by nearest-on-the-network, ⬇ download as a plain-text itinerary (names, addresses, live channels, per-leg distances for both modes, the share link), or share the run through the normal pill row.
• make_plan_demo.py draws the 200×200 looping demo beside the tool by routing on the same graph the page routes on, and writes assets/plan-demo.json so the caption always quotes the run being drawn.

Two networks, not one route at two speeds

Nobody can fly. There are buildings in the way, sois that do not join up, a one-way ring around the moat, and water you cross at a footbridge or a U-turn and nowhere else — and a footbridge is no use to a scooter. So the planner routes on the real network, twice, and a leg usually comes back with two different distances. Straight-line distance is wrong in a city, and most wrong for exactly the two people this page is for.

Pipeline:

python3 importers/crawl_roads.py        # 9 gentle Overpass tiles, ~3 MB cached
python3 importers/build_road_graph.py   # → data/road_graph.json

crawl_roads.py covers the old city plus roughly a 2 km ring (5.5 × 5.5 km), tiled because one box of every highway times Overpass out, and snapshot-first like every other crawl here. build_road_graph.py contracts it to junctions — 9,767 of them, 12,280 edges, 0.56 MB (178 KB gzipped) — keeping road shape for drawing. Only plan.html fetches it.

Every edge carries four permission bits: 1 foot forward, 2 foot back, 4 ride forward, 8 ride back. Three things then make the two modes diverge, all of them tagged in OSM rather than guessed:

• walk-only edges (2,243) — footways, steps, the footbridges over the moat
• edges no walker may use (104) — the flyovers, tagged foot=no
• oneway binds ride and not foot (2,456 vs 11) — this is the one that does the work. On a one-way ring road the shop thirty metres behind you is a lap away, which is precisely why crossing the moat costs a scooter a U-turn and a walker a footbridge.

Worked example, the demo on the page: one leg is 419 m as the crow flies, 526 m on foot, 762 m on a scooter. tests/test_routing.py re-runs the same directed Dijkstra in Python and holds the port to the same answers.

Outside the box nothing is invented — the leg says it is a straight line and offers an OpenStreetMap directions link. No traffic, no turn restrictions, and the two speeds are flat assumptions: distances are real, times are arithmetic.

What this replaced, and why it is worth remembering. An earlier version special-cased the moat: it tested each leg against the ring of four แจ่ง corners and priced a crossing round the nearest gate. It was a real improvement over a straight line and it was still wrong — it put that same leg at 878 m against the network's 526 m on foot. Two lessons kept: a correction that only knows one obstacle will confidently mis-price everything else, and the moat ring still earns its keep as a landmark (tests/test_moat_geometry.py) even though it no longer decides a distance. Also worth keeping: threading the gates into that ring drew a prettier line and broke the containment test, because gate nodes sit on the road crossing slightly outside the water.

Bot hospitality

robots.txt explicitly allows the wildcard and every named AI crawler (GPTBot, ClaudeBot, PerplexityBot, etc) — on purpose, unlike sites elsewhere in this operator's corpus that block them. sitemap.xml lists every page; llms.txt points agents at the open data (data/places.json is the full record dump, data/index.json the slim search index, data/*.geojson per category); every place page also carries schema.org JSON-LD.

Festivals layer

Two tables, not one. A festival is the recurring canon — it carries a date rule, not a date. An event is one dated instance of it, in one year, at one venue. data/festivals.json is the canon (33 entries, hand-curated, confidence marked per entry); data/events.json is the instances. The canon is written once, so only instances need a crawl — which is why 33 festival pages exist without a crawler behind them yet.

festivals_layer.py renders it: /festivals.html (the hub, with the year wheel drawn at build time), a page per festival at /festivals/<id>.html, the standalone /festival-wheel.svg, and /festivals.ics — fixed-date festivals only, with FREQ=YEARLY, because a lunar festival has no date to publish and does not get invented one. Festival venues are resolved to catalogue records through the same strict match_venue the events layer uses, so a wat page carries the festivals it hosts.

build.py calls it in two lines after build_festivals_page(). If those lines go missing, python3 festivals_layer.py re-lays the whole layer over an already-built docs/; --wheel redraws just the wheel.

Getting this year's dates

importers/harvest_festivals.py turns "usually late May" into "26 May – 2 June, announced by the province, here is the page". It reads the provincial PR offices and Chiang Mai municipality — the spec's first-choice sources, TAT and the Chiang Mai PAO, both refuse a plain fetch, and chiangmai.go.th serves a self-signed certificate, all recorded in data/sources.json.

It will not publish a date on its own authority. A row only reaches the site as announced if a canon festival matched, a date parsed, the source is official, the date is in the future, the span is under 45 days, and the month is one the canon already says this festival falls in — and the headline was announcing rather than reporting. Thai government news is overwhelmingly retrospective, and the first live run proved the point: without those gates it produced two dates, both wrong (a marketing slogan with stray digits, and a King's-Birthday merit ceremony matched to Khao Phansa). Everything else is held in data/festival_dates.json as a candidate with the reason it was held, for a human to look at. Nothing on the site reads candidates.

Crawl (gentle by design)

importers/crawl_overpass.py — snapshot-first (cache/overpass/), one query at a time, 12s pauses, retries that rest and rotate mirrors. Refresh with --fetch.

Roadmap

• Phase 2: GIS layer — map pages, ตำบล/ซอย browse tree, landmark-relative "near หอนาฬิกา" queries (condo scouting included).
• Phase 2.5: what's-on feeds — showtimes + events baked like the ticker; weather + horoscope home modules.
• Phase 3: suggest/moderation worker (mueang-map Cloudflare pattern).
• Phase 4: paying sponsors on the advertise.html terms.

Licence

The compilation, the field-collected fields, the prose and the pages: CC BY 4.0. Code: MIT. These are the terms motdang.net/llms.txt and /api/v1/ have been publishing, written down here where a machine cloning the repository will find them.

Attribution string: มดแดง Mot Dang · https://motdang.net/

Anything carried in from elsewhere keeps its own terms, and there is a lot of it — OpenStreetMap under ODbL, Overture Maps Places under CDLA-Permissive 2.0, data.go.th registers under the DGA Open Government License, boundaries under CC BY-IGO, photographs one at a time. Each layer, with its attribution string, is in NOTICE.txt; the verbatim licence text is in LICENSE.

A commercial licence. If attribution does not fit your use — a corpus, a product, a model — write to nan@motdang.net and say what you need.

---

Contact: Nan · nan@motdang.net · Sponsor: Ko-fi · Patreon