Abstract

Modern geospatial systems increasingly require search that is both where-aware and meaning-aware. Traditional spatial indexes (e.g., R-tree, Quad/Oct-tree, S2/Geohash) excel at geometric predicates and topological filtering, yet fall short when users ask semantic questions (“ports similar to Rotterdam,” “neighborhoods with transit-oriented development like X”). In parallel, embedding models for text and imagery enable powerful semantic retrieval but typically ignore spatial topology, containment, and scale.
This paper introduces a hybrid spatial–vector search architecture that unifies spatial predicates with embedding similarity for GIS-scale data. The proposed approach involves: (i) a two-stage retrieval process that initially prunes candidates using spatial cells (such as R-tree or S2 indexing) before ranking results with approximate nearest neighbour (ANN) search over embeddings (for example, HNSW or IVF methods); (ii) cell-aware vector indexes that co-partition embeddings according to space-filling curves, thereby reducing cross-cell probes; (iii) a cost-based query planner designed to jointly optimise spatial selectivity and vector recall; and (iv) a multi-modal Retrieval-Augmented Generation (RAG) layer, which integrates map features, textual data, and remote-sensing image embeddings to produce grounded responses. Evaluation is conducted on public geo-text and satellite imagery datasets, with results reported on latency/recall trade-offs, spatial bias effects, and robustness across heterogeneous scales and coordinate reference systems.
Results demonstrate that hybrid indexing delivers more than tenfold lower latency at fixed recall compared to vector-only baselines for spatially selective queries, while maintaining geometric correctness through predicate pushdown. Integration pathways with mainstream GIS and spatial SQL systems (such as PostGIS combined with pgvector) are explored, and ongoing challenges are identified in areas including geodesic distance metrics, CRS normalization, privacy, and reproducible benchmarking. These findings provide a practical blueprint for AI-empowered geospatial search that addresses both the spatial characteristics of locations and the semantic aspects of meaning.