Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spatially-Enhanced Retrieval-Augmented Generation for Walkability and Urban Discovery

Maddalena Amendola, Chiara Pugliese, Raffaele Perego, Chiara Renso

TL;DR

The paper tackles hallucination and spatial reasoning gaps in LLM-based urban itinerary planning by introducing WalkRAG, a spatial retrieval-augmented generation framework with a conversational interface. It combines a QUAG module for intent-aware dialogue, a Spatial component for walkability-aware routing, and an IR module for grounded information retrieval from a dense-vector index. Experimental results on a Paris dataset show WalkRAG outperforms a closed-book LLM in both spatial task accuracy and information grounding, demonstrating improved walkability scoring and POI integration. This work evidences the viability of spatial RAG to support dynamic, context-aware urban discovery and personalized pedestrian experiences.

Abstract

Large Language Models (LLMs) have become foundational tools in artificial intelligence, supporting a wide range of applications beyond traditional natural language processing, including urban systems and tourist recommendations. However, their tendency to hallucinate and their limitations in spatial retrieval and reasoning are well known, pointing to the need for novel solutions. Retrieval-augmented generation (RAG) has recently emerged as a promising way to enhance LLMs with accurate, domain-specific, and timely information. Spatial RAG extends this approach to tasks involving geographic understanding. In this work, we introduce WalkRAG, a spatial RAG-based framework with a conversational interface for recommending walkable urban itineraries. Users can request routes that meet specific spatial constraints and preferences while interactively retrieving information about the path and points of interest (POIs) along the way. Preliminary results show the effectiveness of combining information retrieval, spatial reasoning, and LLMs to support urban discovery.

Spatially-Enhanced Retrieval-Augmented Generation for Walkability and Urban Discovery

TL;DR

The paper tackles hallucination and spatial reasoning gaps in LLM-based urban itinerary planning by introducing WalkRAG, a spatial retrieval-augmented generation framework with a conversational interface. It combines a QUAG module for intent-aware dialogue, a Spatial component for walkability-aware routing, and an IR module for grounded information retrieval from a dense-vector index. Experimental results on a Paris dataset show WalkRAG outperforms a closed-book LLM in both spatial task accuracy and information grounding, demonstrating improved walkability scoring and POI integration. This work evidences the viability of spatial RAG to support dynamic, context-aware urban discovery and personalized pedestrian experiences.

Abstract

Large Language Models (LLMs) have become foundational tools in artificial intelligence, supporting a wide range of applications beyond traditional natural language processing, including urban systems and tourist recommendations. However, their tendency to hallucinate and their limitations in spatial retrieval and reasoning are well known, pointing to the need for novel solutions. Retrieval-augmented generation (RAG) has recently emerged as a promising way to enhance LLMs with accurate, domain-specific, and timely information. Spatial RAG extends this approach to tasks involving geographic understanding. In this work, we introduce WalkRAG, a spatial RAG-based framework with a conversational interface for recommending walkable urban itineraries. Users can request routes that meet specific spatial constraints and preferences while interactively retrieving information about the path and points of interest (POIs) along the way. Preliminary results show the effectiveness of combining information retrieval, spatial reasoning, and LLMs to support urban discovery.

Paper Structure

This paper contains 8 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. The WalkRAG framework
  3. Query Understanding and Answer Generation
  4. Spatial component
  5. Information Retrieval component
  6. WalkRAG assessment
  7. Results and discussion
  8. Conclusion and future work

Figures (2)

  • Figure 1: The WalkRAG framework. The user query asking for a route from Notre Dame to the Eiffel Tower is redirected by the QUAG component to the spatial component for itinerary construction and walkability score computation (1). The answer returned to QUAG is interpreted by the LLM and returned to the user (2), who further interacts with the conversational system, asking for more details on Champs de Mars (3). QUAG now redirects the query to the Information Retrieval component, which retrieves the appropriate content from an index. The results retrieved are interpreted by the LLM and returned to the user (4).
  • Figure 2: LLM‑CB and WalkRAG routes for the third spatial query of the dataset.