Thai Travel AI

AI Travel Planner สำหรับวางแผนท่องเที่ยวในประเทศไทย โดยเน้น Road Trip

Project Goal

สร้าง AI ที่เข้าใจภาษาไทย ช่วยวางแผนการเดินทาง และทำงานร่วมกับข้อมูลสถานที่จริงได้

Team Responsibilities

• Person 1: AI Engineer
• Person 2: Travel Data & Tools Engineer
• Person 3: Backend & Integration Engineer

AI Engineer Responsibilities

• Design the system prompt
• Connect an AI model through an API
• Design the travel-planning conversation
• Return structured travel-plan data
• Test Thai responses and reduce hallucinations

Initial Milestone

The AI can accept a travel request in Thai and return a structured sample itinerary.

Run locally

1. Put GROQ_API_KEY=... in ai-engine/.env (never commit it), then install both packages:

cd ai-engine && npm install
cd ../services/api && npm install

2. Start the API (port 3000):

cd services/api
npm start

3. Serve the repo root with any static server, then open http://localhost:8080:

python -m http.server 8080

The page calls http://localhost:3000. Override it with window.TRAVEL_API_BASE_URL before script.js loads, or add ?api=http://host:port to the page URL.

When the page is opened from localhost, 127.0.0.1 or a file, it calls the API at http://localhost:3000. On any other host (for example a Vercel URL) it calls /api on the same domain.

Groq's free tier allows about 8,000 tokens per minute, which is roughly one plan per minute. When the quota is full the API answers 503 and the page asks the user to wait.

Deploy to Vercel

The repo deploys as one Vercel project: the root HTML/CSS/JS are served as static files, and every /api/* request goes to one serverless function (api/index.js) that runs the same Express app as services/api (services/api/app.js).

Files involved:

• package.json (repo root): dependencies the function needs (express, cors, dotenv, groq-sdk).
• api/index.js: exports the Express app as the function.
• vercel.json: sends /api/* to the function, bundles data/places/**, places_bkk_cnx.json and ai-engine/system-prompt.md with it, and allows up to 60 s per request.

Steps:

1. Push the branch you want to deploy to GitHub.
2. In Vercel, choose Add New → Project and import tanyimcoffee-hub/AFO_team_fatzone.
3. Keep Framework Preset: Other and the Root Directory as the repo root. Leave the build command and output directory empty.
4. Under Environment Variables, add GROQ_API_KEY with your Groq key. Optional: OSRM_URL to use your own routing server.
5. Click Deploy. To deploy a branch other than main, either set it as the Production Branch in Settings → Git, or open the branch's Preview deployment.
6. Check https://<your-project>.vercel.app/api/health, then open the site and create a plan.

Notes:

• Never put GROQ_API_KEY in the code. If it is missing, /api/plan answers 503 with a clear message.
• Each new function instance loads about 11.5 MB of place data first, so the first request after a quiet period is slower.
• Vercel's Hobby plan caps a function at 60 seconds. AI or routing calls that take longer fail and the page shows a retry button.
• The public OSRM demo server is free but best effort. For heavy use, run your own OSRM server and set OSRM_URL.

To test the same setup locally, install the Vercel CLI and run vercel dev from the repo root (it needs GROQ_API_KEY in a root .env, which is git-ignored).

Web pages

Repo root, sharing styles.css and script.js (the browser script; not to be confused with services/api/app.js):

• index.html (landing): quick-start box (origin, destination, days) that opens the chat page, ready-made trip templates, auto-scrolling feature cards, a region picker for all 77 provinces and a per-province place explorer.
• plan.html (chat page): a large chat window (trip form, AI replies, reply box for follow-up questions) next to a large map. Links can prefill it, e.g. plan.html?origin=กรุงเทพมหานคร&destination=จันทบุรี&days=2&go=1 (go=1 sends to the AI right away; other params: budget, travelers, tripType, interest).
• The map (Leaflet + OpenStreetMap tiles) draws the plan's stops and the driving route along real roads from /api/route, with total distance and driving time. If routing fails, it falls back to a dashed straight line and says so. If Leaflet cannot load, an illustrative SVG diagram is shown instead.

API (services/api)

Method	Path	Notes
GET	/api/health	Status check
POST	/api/plan	Body: origin, destination (required), days, budgetTHB, interests[], tripType (one_way/round_trip), travelers (1–50), message (follow-up text, ≤1000 chars), history (≤20 × {role: user|assistant, content}). Returns data = AI response (reply, status, trip, itinerary; stops get lat/lon when known) and meta (places sent, provinces, approximate origin/destination points, attribution).
GET	/api/places	Query: province (Thai or English name), type, q, limit (1–200, default 50), offset.
GET	/api/places/provinces	All 77 provinces with counts per type.
POST	/api/route	Body: points (2–25 × {lat, lon} in driving order). Returns distanceKm, durationMin, legs, geometry ([lat, lon] list) and attribution. Uses the public OSRM demo server by default (free, no key, best effort); set OSRM_URL to use another OSRM server. 400 bad input, 422 no road route, 502 routing service unavailable.

Errors are { success: false, error } with 400 (bad input), 404 (unknown province), 502/503 (AI quota or bad AI output) or 500.

Tests: npm test in services/api (store) and in ai-engine (validators).

Place data

See data/places/README.md. Starter data comes from OpenStreetMap (© OpenStreetMap contributors, ODbL) and is marked unverified. Travel Data team files in data/places/team/ (and the original places_bkk_cnx.json) override OSM rows.