Case study
Agentic RAG for Permit and Property Intelligence
Property questions answered with records, sources, and map context.
The challengeManual searches across fragmented permit and property datasets.
The solutionConversational retrieval with source-linked answers and map pins.
Semantic retrieval connected to spatial and property lookup.
Cited answers, map pins, and exportable results.
Approximately 70 engagements per day reported.
Engagement
Overview
Permit Signal makes permit and property information easier to discover through a chat-and-map experience. The system combines the meaning of a user’s question with location and property context to retrieve relevant records.
I built the data and retrieval workflow around an ingestion scope of 3 million New Jersey construction permits and 60,000 property records from Socrata and ArcGIS. The pipeline normalizes and geocodes the data into PostGIS, connects records through spatial joins, and prepares searchable representations.
Behind the Signal conversational agent, the hybrid retriever combines spatial lookup, semantic search, BM25 keyword retrieval, and SQL queries. Retrieved evidence and property coordinates feed source-linked answer assembly, with a language-model endpoint supporting response wording. Users receive answers alongside map context and can export results.
Usage figures are internally reported. Reporting periods and user definitions remain unconfirmed. Engagement definitions also need confirmation; the dataset counts describe documented ingestion scope, not a coverage or freshness guarantee.
- Reported users
- 1,797
- Reported engagements / day
- ≈70
- Reported messages / day
- ≈480
- Documented permit ingestion scope
- 3M
- Documented property ingestion scope
- 60K
Challenge
Useful property answers needed both relevant information and the location context connecting the records.
Fragmented data
Permit and property sources needed normalization before they could support a combined search experience.
Spatial relationships
Geocoding and joins were needed to connect records to properties and locations.
Different query types
The workflow needed semantic search alongside location filters, keyword matching, and structured queries.
Traceable results
Answers needed supporting sources and map context, not only a generated summary.
Solution
I connected the data pipeline, retrieval methods, and map experience around the user’s question.
Data ingestion
Bring together permit and property records from Socrata and ArcGIS.
Geospatial preparation
Normalize, geocode, and join records using PostGIS.
Hybrid retrieval
Combine spatial lookup, semantic search, BM25, and SQL.
Conversational orchestration
Connect the agent’s semantic queries and location/property filters to the retrieval layer.
Source-linked generation
Assemble responses from retrieved evidence and relevant coordinates.
Map and export outputs
Return map pins and make results available for export.
Why hybrid retrieval matters
Permit and property questions can depend on both text and location. This workflow combines semantic retrieval with spatial and structured lookup so the answer is supported by relevant records and property context. The language model helps express the result; the retrieval system supplies the evidence behind it.
Impact
Reported usage shows recurring adoption of a source-linked discovery workflow over the documented permit and property dataset.
Reported adoption
1,797 users.
Reported engagement
Approximately 70 engagements per day.
Reported conversation activity
Approximately 480 messages per day.
Implemented capability
Conversational retrieval with cited answers, spatial context, map pins, and exportable results.