1sdv/TripStar

TripStar – AI‑powered travel planning assistant

What it does

• TripStar is a web application that builds complete travel itineraries for a user‑specified trip (cities, dates, preferences). It pulls data from real‑world sources (Xiaohongshu travel posts, weather services, hotel listings, map APIs) and uses a large language model (LLM) together with a set of specialized agents to turn that raw data into a structured travel plan.

How it works

1. Front‑end (Vue 3) – A form lets users enter destination, travel dates and personal preferences (transport mode, accommodation style, interests, etc.). The UI shows progress via WebSocket updates and, once the plan is ready, displays:
   • Day‑by‑day schedule with attractions, travel routes, hotel suggestions, meals, and weather forecasts.
   • Budget breakdown (tickets, food, transport, accommodation).
   • Interactive maps (Google Maps for overseas, AMap for China) with routes drawn between points.
   • A knowledge‑graph visualisation of the itinerary.
   • A floating AI chat window that can answer follow‑up questions using the full itinerary as context.
2. Back‑end (FastAPI, Python 3.10+) – The request /api/trip/plan starts an asynchronous task and returns a task_id. A background worker runs a multi‑agent workflow:
   • Xiaohongshu Agent searches Xiaohongshu (a Chinese social platform) for travel posts, scrapes the content, and asks the LLM to extract a JSON list of attractions, their typical visit duration, and whether they need advance reservation.
   • Weather Agent fetches forecast data via Google or AMap weather APIs.
   • Hotel Agent queries POI services for hotels that match the budget and style.
   • The Planner Agent combines all pieces, builds an optimal route (avoiding back‑tracking), and asks the LLM to output a detailed itinerary plus a budget in a strict JSON format.
   • The result is parsed, cleaned, and fed into a knowledge‑graph service that creates nodes/edges for cities, days, attractions, and costs.
3. Data enrichment – After the itinerary is generated, the front‑end requests photos for each attraction. The back‑end again calls Xiaohongshu, grabs the first image from a recent post, and returns a direct URL so the UI shows real‑world photos.

Key technical highlights

• Multi‑agent architecture built on the HelloAgents framework, with clear role separation (attraction extraction, weather, hotels, routing).
• LLM‑driven extraction: long, unstructured travel notes are distilled into clean JSON by prompting the model.
• Dual map engine: automatic fallback between Google Maps (outside China) and AMap (inside China) for geocoding, routing and map rendering.
• Async task queue + WebSocket progress updates to avoid 504 gateway timeouts during long LLM calls.
• Internationalisation: UI and LLM prompts support Chinese, English and Japanese; generated itineraries are automatically translated.
• Budget panel and knowledge‑graph visualisation give users a financial overview and a network view of their trip.
• Docker‑compose deployment with environment‑variable configuration for LLM API keys, map service keys, and Xiaohongshu cookies.

Who might use it

• Travelers who want a personalized, end‑to‑end itinerary without manually browsing multiple sites.
• Developers interested in building multi‑agent LLM applications that combine web‑scraping, structured data extraction, and real‑time UI updates.

Getting started

1. Install Python 3.10+, Node 18+, and Docker (optional).
2. Provide API keys for an OpenAI‑compatible LLM, Google Maps, AMap, and a logged‑in Xiaohongshu cookie.
3. Run the backend (uvicorn app.api.main:app) and the Vue front‑end (npm run dev) or launch both via docker‑compose up.
4. Open the web UI, fill in the trip details, and watch the system generate a complete plan.

---

Bottom line: TripStar is a full‑stack, multi‑agent AI system that turns raw travel data into a polished, budget‑aware itinerary, complete with maps, photos, and an interactive AI assistant.