kth.sePublications KTH
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Grounding agentic AI in building semantics: From information retrieval to automated simulation
KTH, School of Architecture and the Built Environment (ABE), Civil and Architectural Engineering, Building Technology and Design.ORCID iD: 0000-0003-1509-0195
KTH, School of Architecture and the Built Environment (ABE), Civil and Architectural Engineering, Building Technology and Design.ORCID iD: 0000-0001-5916-7890
RISE Research Institutes of Sweden, Division of Digital Systems, Computer Science, Isafjordsgatan 22, Kista, 164 40, Sweden.ORCID iD: 0000-0001-5091-6285
KTH, School of Architecture and the Built Environment (ABE), Civil and Architectural Engineering, Building Technology and Design.ORCID iD: 0000-0001-6266-8485
2026 (English)In: Energy and Buildings, ISSN 0378-7788, E-ISSN 1872-6178, Vol. 371, article id 118187Article in journal (Refereed) Published
Abstract [en]

Building intelligence remains limited by fragmented data, labor-intensive semantic modeling, and AI interfaces that lack grounding in verifiable building facts. This study presents an Agentic AI framework that combines a simplified semantic modeling approach, termed KebGraph, with large language model (LLM)-based AI agents. The framework enables end-to-end workflows from natural language instructions to engineering services. KebGraph organizes heterogeneous building context into a queryable semantic layer, while a multi-agent architecture maps user requests to concrete engineering actions through tool execution. Evaluated on a real office building, the framework is tested across four large language models on a 213-question benchmark. Grounding on KebGraph attains the highest factual question-answering accuracy of the compared representations, ahead of a conventional Brick model and a raw-document baseline, while using the fewest input tokens among the grounded representations. In an agentic function-calling evaluation, it also achieves the highest average argument-formation score across the evaluated models. On invalid or adversarial requests, the capable models largely decline or ask for clarification rather than retrieving fabricated data. An indoor thermal modeling workflow further demonstrates agent-driven construction, calibration, and evaluation of thermal models end to end over a one-month winter period. These results indicate that grounding LLM agents in a lightweight building semantic layer improves the accuracy, efficiency, and reliability of information retrieval and tool use on this building, while broader generalization across buildings and seasons remains future work.

Place, publisher, year, edition, pages
Elsevier BV , 2026. Vol. 371, article id 118187
Keywords [en]
Agentic AI, Building semantics, Indoor thermal simulation, Knowledge graph, Large language model agents
National Category
Building Technologies
Identifiers
URN: urn:nbn:se:kth:diva-388712DOI: 10.1016/j.enbuild.2026.118187ISI: 001876625900001Scopus ID: 2-s2.0-105049664480OAI: oai:DiVA.org:kth-388712DiVA, id: diva2:2104703
Note

QC 20260924

Available from: 2026-09-24 Created: 2026-09-24 Last updated: 2026-09-24Bibliographically approved

Open Access in DiVA

No full text in DiVA

Other links

Publisher's full textScopus

Authority records

Peng, ZengChen, YangzheWang, Qian

Search in DiVA

By author/editor
Peng, ZengChen, YangzheOhlson Timoudas, ThomasWang, Qian
By organisation
Building Technology and Design
In the same journal
Energy and Buildings
Building Technologies

Search outside of DiVA

GoogleGoogle Scholar

doi
urn-nbn

Altmetric score

doi
urn-nbn
Total: 26 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf