GeoQuery | Gerard Calvo Bartra

GeoQuery transforms geographic data analysis by combining Large Language Models with spatial databases. Simply ask questions in natural language and get instant maps, charts, and insights.

PROJECT

GeoQuery

IMPACT AREAS

generative-ai
data-viz

SERVICES

Python
React
DuckDB

Challenge: Democratizing access to geospatial analysis

Geographic data analysis traditionally requires specialized GIS knowledge, complex SQL queries, or expensive proprietary software. This creates a significant barrier for urban planners, researchers, and decision-makers who need spatial insights but lack technical expertise.

The challenge was to create a system that could understand natural language questions about geographic data and automatically generate accurate spatial queries, visualizations, and explanations — making territorial intelligence accessible to everyone.

Approach: LLM-powered spatial query pipeline

GeoQuery combines Large Language Models with spatial databases through a multi-stage pipeline. The system processes natural language queries through intent detection, semantic dataset discovery, and automated SQL generation.

Intent Detection — LLM classifies the user's query type (map request, chart, comparison)
Semantic Search — Embedding-based discovery finds relevant datasets from 100+ options
SQL Generation — LLM creates spatial SQL queries using DuckDB Spatial syntax
Execution & Visualization — Results are automatically styled as maps, charts, or tables

GeoQuery interface showing natural language query and map visualization — GeoQuery transforms natural language questions into interactive map visualizations with automatic styling.

Key Features: Natural language meets spatial intelligence

🗣️ Conversational Queries — Ask questions in plain English instead of writing SQL
🗺️ Auto-Visualization — Smart choropleth maps, point markers, and heatmaps generated automatically
🔍 Semantic Discovery — AI embeddings find relevant datasets from 100+ options
🧩 Multi-Step Analysis — Complex queries automatically decomposed and executed
💡 Thinking Transparency — See the LLM's reasoning process streamed in real-time

Technical Details: Modern full-stack geospatial architecture

Backend built with Python 3.11 and FastAPI with SSE streaming
DuckDB Spatial for in-memory spatial analytics and SQL execution
Google Gemini 2.0 for intent detection, SQL generation, and explanations
Next.js 15 frontend with Leaflet for interactive web maps
Sentence-transformers for semantic dataset search via embeddings
100+ datasets including administrative boundaries, infrastructure, and socioeconomic indicators

Impact: Territorial intelligence for everyone

GeoQuery demonstrates that natural language interfaces can make powerful spatial analysis tools accessible to non-technical users. By abstracting away SQL complexity and GIS software learning curves, the platform enables urban planners, researchers, and policymakers to extract insights from geographic data through simple conversation.

The open-source architecture provides a template for building similar territorial intelligence platforms for any region, with modular components that can be adapted for different datasets, languages, or domain-specific applications.

Technologies

This project was built with:

↗ Python

↗ React

↗ DuckDB