UrbanRAG

UrbanRAG is a spatially-enhanced Retrieval-Augmented Generation (RAG) framework for recommending urban itineraries. It integrates Large Language Models (LLMs) with spatial reasoning and geographic data to support exploratory itinerary and urban discovery through a conversational interface.

This project is associated with the paper:

Spatial RAG for Big Data–Driven Urban Itinerary Recommendation
Maddalena Amendola*, Chiara Pugliese*, Raffaele Perego, Chiara Renso
*Both authors contributed equally to this research.

---

🚶 What is UrbanRAG?

UrbanRAG helps users explore cities by generating personalized routes that prioritize:

• Domain-specific spatial indicators (e.g., sidewalks, air quality, greenery, accessibility)
• User preferences (e.g., types of POIs or route aesthetics)
• Contextual information (e.g., fun facts, history, or details about places encountered along the route)

The system combines three main components:

• QUAG: Query understanding and answer generation via an LLM (Llama 3.1 8B)
• Spatial Component: Computes routes and scores them based on domain-specific spatial indicators
• IR Component: Retrieves relevant context passages using a dense vector index (FAISS + Snowflake bi-encoder)

---

🔍 Key Features

• Natural language queries for both routes and urban information
• Multi-criteria assessment using OpenStreetMap and environmental data
• Retrieval-augmented generation to mitigate hallucinations and improve accuracy
• Custom dataset for evaluating spatial and information-seeking queries

---

📂 Repository Structure

• main.py – sets up and runs the interactive UrbanRAG system
• QUAG.py – LLM-based query classification and generation
• src/spatial_component/ – Route generation, itinerary scoring, and POI enrichment
• src/RAG_system/ – Dense passage indexing and neural search
• dataset/ – UrbanRAG evaluation dataset (10 route queries + 30 follow-ups)
• output/ – UrbanRAG outputs

---

📊 Evaluation Summary

Query Type	Metric	UrbanRAG	LLM-Only
Spatial Queries	Fully Correct	4 / 10	0 / 10
	Partially Corr	6 / 10	0 / 10
Info Queries	Fully Correct	20 / 30	12 / 30
	Partially Corr	5 / 30	11 / 30

UrbanRAG clearly outperforms closed-book LLMs in both spatial understanding and factual accuracy.