HIPAT Persona and Floor Plan Topology Explorations

This repository contains two exploratory AI workflows developed in FT-2 Künstliche Intelligenz of the NAH AM NUTZEN 2 (NAN-2) project:

1. a HIPAT persona for privacy-oriented design review;
2. a floor plan topology pipeline for extracting spaces, adjacency, and direct connectivity from floor plan images.

The topology experiments informed the later EG-ICE 2026 study A Multimodal LLM-Based Approach to Generating Topology Graphs of Buildings from Floor Plan Images. The full paper pipeline and benchmark are maintained separately in design-computation/eg-ice-2026-mllm-topology-graph.

Repository contents

• notebooks/hipat_persona.ipynb — extracts a structured spatial program and produces privacy-related risks and design recommendations using a simplified HIPAT-based evaluation.
• notebooks/floorplan_topology_pipeline.ipynb — extracts labeled spaces and their spatial relationships from a floor plan image and compares the result with a reference graph.
• notebooks/floorplan_topology_analysis.ipynb — summarizes the included topology experiment results.
• data/resplan_2769_mono_room_annot.png — example floor plan from ResPlan.
• data/resplan_2769_reference_graph.md — manually authored reference graph for the example plan.
• results/floorplan_topology_metrics.json — summarized results from repeated topology-extraction runs on the example plan.
• results/privacy_planning_persona_examples.md — the full prompts and evaluator responses before and after operating context was added for three compact two-bed room variants.
• environment.yml — Conda environment definition.
• .env.example — example configuration for supported model providers.

HIPAT persona

The HIPAT notebook explores how an LLM-based persona can support privacy-oriented evaluation of architectural ideas. It converts design information into a structured description of spaces, users, occupancy, access, and spatial relationships and then generates privacy-related risks and recommendations.

The evaluation uses a simplified mapping derived from the six-level Hierarchy of Isolation and Privacy in Architecture Tool (HIPAT). The prototype was developed for small design cases in NAN-2, including double-room variants from the CAFÉ PLATTFORM project. Operational information such as supervision, ventilation, room-use rules, and restrictions on movable partitions was included because the spatial geometry alone was not sufficient for meaningful recommendations.

This is an exploratory adaptation of HIPAT developed for the specific use case in this repository. It uses a simplified mapping of privacy levels based on users, occupancy, and access.

Reference:

McCartney, Shelagh, and Ximena Rosenvasser. 2022. “Privacy Territories in Student University Housing Design: Introduction of the Hierarchy of Isolation and Privacy in Architecture Tool (HIPAT).” SAGE Open 12(2). DOI

Floor plan topology extraction

The topology workflow tests whether a multimodal language model can derive a simple graph representation directly from a floor plan image.

The graph distinguishes two spatial relationships:

• Adjacency — two labeled spaces share a boundary.
• Direct connectivity — direct traversal between two spaces is possible through a door, opening, or passage.

The public example uses one cleared ResPlan floor plan and a manually authored reference graph. Repeated model runs are compared with this reference to identify missing and unexpected relationships.

This is a single-plan exploratory prototype. The included metrics are not the benchmark results reported in the EG-ICE paper. The later study substantially extends this work using a reviewed 100-plan benchmark derived from Modified Swiss Dwellings (MSD), prompt optimization, validation and repair, repeated runs, and cross-model evaluation.

Setup

Create and activate the Conda environment:

conda env create -f environment.yml
conda activate hipat-research-artifact

Copy the example environment configuration and add credentials for the model provider you want to use:

cp .env.example .env

Then start Jupyter from the repository root:

jupyter lab

The notebooks support the provider configurations documented in .env.example. Do not commit .env or API credentials.

Model-backed notebooks may produce different outputs across runs and may incur API charges. The analysis notebook can be used without model credentials.

Data and limitations

Only the cleared ResPlan example included under data/ is published here. Project-specific CAFÉ PLATTFORM drawings and other development inputs are not part of the public repository. The persona examples use text-only room descriptions and selected evaluator responses.

Licenses

Code is licensed under the MIT License. The cleared data, reference graph, summarized results, and documentation are licensed under CC BY 4.0. Third-party datasets and model services remain subject to their own licenses and terms.