Advanced RAG Engineering for real estate due diligence AI Agent.

The challenge

Municipal PDFs, parcel GIS, and models that invent a zoning class

The agent had to run due diligence across US municipalities. That meant stitching high-volume streams — geospatial, environmental, municipal, legal, infrastructural — into insight-based reports that stay accurate. Parcel variables include land use, zoning class, total acres, proposed use and zoning, open space, greenways, wetlands, surface water, watersheds, adjacent and planned roads, conservation elements, trails, and access easements.

Generic tools could not follow Mapline.AI's workflow or the variance in those sets. They hallucinated. Municipal regulation differences made a custom agent the only honest path.

How it works

Four RAG moves past a vanilla retrieve-then-generate

The agent analyses legal, environmental, and planning documents and writes report sections. Basic RAG was not enough. Four custom layers raise accuracy and relevance.

Parsing complex documents with specialized OCR

Real estate PDFs mix tables, images, and awkward layouts. Standard extraction fails. Mistral OCR looked strong on public benchmarks and consistently underperformed on Mapline.AI's hardest files. LlamaParse Premium parsed those tables. Clean text into the LLM is the difference between a usable report and a fluent error.

Wojciech Achtelik, AI Lead Engineer at Vstorm: “Once we understood the challenge, we chose to implement a more robust OCR solution. Standard text extraction tools struggled with the complexity of real estate documents, particularly those with intricate tables. By adopting a specialized OCR setup, we achieved more accurate parsing of these challenging documents, ensuring that LLMs receive clean, reliable text for analysis.”

Metadata filtering in Milvus

The retrieval store is Milvus because it can filter on metadata. Documents are tagged with municipality and type. A query can restrict to that location and type, so the model sees the right file, not the nearest-looking file from another town.

Custom re-ranking

Not every hit is equal. After the first retrieve, a re-ranker boosts sources known to be authoritative for that report section.

Dynamic geospatial extraction

Structured parcel facts — zoning codes, land use, adjacent roads, nearby wetlands — are pulled from geospatial data and injected into the prompt with the document text. The model is not guessing those fields from prose alone.

Results

What Mapline.AI can sell

The shipped agent is specific to real estate due diligence, not a chatbot with a property theme. Diligence that took weeks can run in minutes. An end user put potential savings at $2,000–$3,000 on a 90-acre parcel that still needed a sketch plan and consultants. The system also surfaces zoning quirks and environmental details that a rushed manual review often skips.

LangSmith watches the agent in production. Encryption and compliance steps sit on the sensitive data. The architecture is meant to take more volume as Mapline.AI's corpus grows. Mapline.AI's team updates zoning and property records themselves. Feedback from that team and from users continues to tune accuracy.