YelpAI Set Go!

More than a simple agent, YelpAI-Set-Go is an AI travel companion!

Comment

Inspiration

Trip planning usually happens across too many tabs: “where should we eat?”, “what’s nearby?”, “can we get a table?”, “how far is that from where we are?”. We wanted to compress that whole workflow into a single conversation—powered by real business data—so you can plan a great trip without doing the busywork.

What it does

  • Lets you chat with a primary travel agent (YelpAI-Set-Go!) to describe your trip, preferences, dates, and constraints.
  • Uses Yelp AI (clientId: pN6pzJR6mK7549M0azoaxg) to discover real restaurants/attractions/services and summarize recommendations.
  • Generates a structured, day-by-day, hour-by-hour travel itinerary and displays it in an itinerary dashboard.
  • Saves itineraries in the browser and lets you reload a prior itinerary (optionally with its chat history).
  • Shows driving directions on an embedded map from your current location to a selected itinerary stop.
  • Supports Yelp bookings workflows (openings, holds, reservations, status, cancel) via tool calls.

How we built it

  • UI: Blazor Server (.NET 10) with interactive server components and SignalR streaming.
  • Agent orchestration: Microsoft.Agents.AI + Microsoft.Extensions.AI with tool/function calling.
  • LLM provider: OpenRouter via the OpenAI .NET client configured with the OpenRouter endpoint.
    • Primary agent defaults to openai/gpt-5.1.
    • Itinerary generation uses a separate agent with JSON-schema output (e.g., openai/gpt-oss-120b).
  • Data + actions exposed to the agent as tools:
    • Yelp AI Chat API for discovery (https://api.yelp.com/ai/chat/v2).
    • Yelp Bookings endpoints for reservation workflows.
    • Weather via Open-Meteo.
    • Itinerary creation/modification tools.
  • Location + maps:
    • Browser geolocation + reverse geocoding (Nominatim / OpenStreetMap) to turn coordinates into a human-readable starting point.
    • Google Maps JavaScript API for directions + markers, called via JS interop.
  • Persistence:
    • Itineraries and serialized agent threads are stored in browser local storage so you can come back later.

Challenges we ran into

  • Streaming UI: keeping the chat responsive while tokens stream and avoiding excessive re-rendering.
  • Tool inputs/outputs: Yelp AI is multi-turn. Tool input descriptions needed to be very specific about when to include the chat_id.
  • Multi-turn context: persisting/rehydrating agent threads and associating them with saved itineraries.
  • Maps integration: coordinating JS interop, DOM lifecycle, and map updates as destinations change.

Accomplishments that we're proud of

  • A clean “Chat / Itinerary” workflow that feels like a single experience, not separate pages.
  • Real place grounding: recommendations come from Yelp APIs, not hallucinated venues, real-time maps and actual weather forecasts.
  • One-click return trips: saved itineraries (and optional chat history) load back into the session.
  • In-context navigation: map directions pop up directly from itinerary selections.

What we learned

  • Tool calling becomes dramatically more useful when the UI reflects tool outcomes (itinerary cards, maps, saved state).
  • Persisting agent threads makes the experience feel continuous—but requires careful UX around “clear session” vs “load history”.
  • Blazor Server + streaming is a strong fit for conversational UX when you throttle renders intentionally.

What's next for YelpAI Set Go!

  • Add a first-class “edit itinerary” flow (inline swaps, drag/drop, and smarter constraints).
  • Improve booking UX (confirmations, error recovery, and clearer reservation summaries).
  • Expand map modes (walking/transit, multi-stop routing) and deeper itinerary-to-map visualization.
  • More robust persistence in an external DB (browser storage is obviously demo-only).

Built With

Share this project:

Updates