✈️ Roamie — Multi-Agent Travel Planner

A proof-of-concept travel assistant built from cooperating AI agents — each an expert in one part of the trip.

Roamie explores multi-agent orchestration: instead of one monolithic model, the trip is split across specialized agents that each own a domain — sightseeing, dining, and flights — and expose their skills as callable tools. It's built on smolagents with HuggingFace inference models.

---

The Agents

Agent	File	What it does
🗺️ Travel Guide	travel_guide.py	Recommends activities and sights based on destination + interests (history / food / adventure)
🍽️ Restaurant Guide	restaurant.py	Suggests notable restaurants for a given city
🛫 Flight Search	flight_booking.py	Parses a natural-language flight query and returns matching options

Each agent is a CodeAgent with a single focused @tool, designed so its string output can be consumed by another agent — the foundation for chaining them into a full itinerary planner.

---

Notable Engineering

The flight agent shows the most depth:

• LLM-first parsing with graceful fallback — tries a HuggingFace model to extract origin, destination, dates, and passengers from free text; if no API key or the call fails, falls back to a regex parser so it always works offline
• Tool-call adapter (call_tool) that transparently handles LangChain StructuredTool, .func, .run, or plain callables
• Structured JSON results ready for downstream agents

---

Setup

pip install smolagents langchain requests

# Optional — enables LLM-powered query parsing & recommendations
export HF_API_KEY="your_huggingface_token"
export HF_MODEL="google/flan-t5-small"   # optional override

Run

# Flight search demo
python flight_booking.py

# Or call an agent directly
python -c "from travel_guide import travel_guide_bot; print(travel_guide_bot('Things to do in Tokyo, I love history'))"

---

Status & Roadmap

This is an early prototype — recommendation data is currently curated/mocked while the agent-orchestration layer is built out.

• Orchestrator agent that chains guide → restaurant → flights into one itinerary
• Replace mocked recommendation tables with live APIs (Places, flight search)
• Conversational memory across the planning session
• Web UI