Every US county has been encoded into a graph database connecting hazard risks, social vulnerability, economic hardship, health outcomes, and disaster declaration history. Instead of searching spreadsheets, you ask a question and get a synthesized answer grounded in that data.
FEMA NRI, CDC SVI, CDC PLACES, Census ACS, FEMA Disaster Declarations — one structured document per county
18 hazard types per county: wildfire, hurricane, flood, tornado, earthquake, drought, hail, ice storm, strong wind, and more
Extracts entities and relationships. Builds the graph: county → hazard → risk index → health indicator → declaration history. Running live at explorer.jbf.com
Plain English question → graph traversal + vector search → synthesized answer with citations. Any app can call the API.
Each county document contains structured data from five federal sources: FEMA National Risk Index (risk scores for 18 hazard types), CDC Social Vulnerability Index, CDC PLACES health outcomes, Census ACS socioeconomic indicators, and FEMA Major Disaster Declaration history. LightRAG ingests these as text and extracts entities and relationships automatically using an LLM.
LightRAG reads each document and extracts named entities (county names, hazard types, index names) and the relationships between them. These become nodes and edges in a graph database. When you see Terrebonne → Hurricane in the graph, that's a stored relationship with associated risk data. "Degree 19" means a county has 19 direct connections — one for each hazard type and risk index it's linked to.
When you ask a question, LightRAG runs two searches simultaneously: (1) traverses the graph to find relevant entities and their neighbors, and (2) runs vector similarity search over the raw document chunks. Combining both results is what allows it to answer specific factual questions ("what is Terrebonne Parish's hurricane risk?") and complex comparative questions ("which Louisiana parishes have both high flood risk and low healthcare access?").
The retrieved graph context is passed to Claude with your question. The model synthesizes a natural language answer from the graph data — not from general training knowledge. This is Retrieval Augmented Generation (RAG): the AI is grounded in your specific dataset. Answers include citations back to source counties and federal datasets.
LightRAG exposes a REST API. Any application can send a natural language question and receive a structured answer. The next evolution is Spatial RAG: connecting this graph to ArcGIS maps so a disaster planner can click a county on a map and ask "what's this county's compounded risk profile?" — answered in seconds by the knowledge graph, displayed directly in the map interface.