UNSW Postdoctoral Researcher Dr Rabindra Lamsal, together with Prof. Sisi Zlatanova, recently presented an update on their research project,

“Geospatial Intelligence to Support Machine–Human Native Conversation”, at the FrontierSI staff meeting.

The project aims to design a conversational system that enables natural language interaction with complex geospatial datasets. It is structured in two phases:
(i) developing a conversational system for Industry Foundation Classes (IFC) models; and
(ii) extending this capability to support combined 3D city models and IFC models.

The presentation focused on the completion of the first phase.

Working with IFC models typically requires substantial technical expertise, including proficiency in specialised BIM tools, software libraries and programming environments. These barriers often limit meaningful engagement with IFC data to technical specialists. While Large Language Models (LLMs) present new opportunities for natural language interaction, their limited context capacity — combined with the large size and structural complexity of raw IFC files — makes it impractical to directly ingest entire IFC datasets for querying.

To address this challenge, the research team developed a hybrid relational–graph LLM framework for natural language querying of IFC models. Rather than relying on raw file input, the framework transforms IFC data into structured relational and graph database representations that are more efficient for LLM-based traversal and reasoning. This architectural approach enables scalable, accurate querying without exceeding model limitations.

An implementation of the framework was demonstrated during the presentation. Testing across three IFC models of varying sizes achieved accuracy rates between 93.3% and 100%, demonstrating both robustness and adaptability.

The hybrid framework represents the principal contribution of Phase One and the associated manuscript is currently under journal review. In parallel, the team has developed a complementary semantic retrieval helper system for the IFC schema, designed to enhance performance in similar LLM-based querying frameworks. This supporting work has been accepted for publication in the *Annals of the ISPRS Congress 2026*.

The session provided FrontierSI staff with a comprehensive overview of progress to date and set the foundation for Phase Two of the project, which will extend the conversational system to integrated 3D CityGML and IFC environments.