Supplemental Materials for Benchmarking Retrieval-Augmented Large Language Models for Green Building Cost and Duration Prediction

Abstract

Supplemental materials for a study benchmarking retrieval-augmented large language models for predicting green building project cost and duration. Forty-one configurations from Anthropic, Google and OpenAI, across their reasoning-effort settings, each predicted 20 held-out Hong Kong BEAM Plus projects three times (2,460 predictions), grounded in the ten most similar of 79 historical projects retrieved through a Qdrant hybrid dense-and-sparse pipeline. Contents:S1 model identifiers and reasoning levels, all 41 configurationsS2 excluded configurations, with reasonsS3 rank stability across 66 weightings, all 41S4 full leaderboard, all 41S5 pairwise statistical comparisons against the best configurationS6 classification accuracy, top configurationsS7 ratio-target error metrics, all 41S8 mean tokens per prediction, all 41S9 reproduction check on a rebuilt index, best cost-MAPE configurationS10 magnitude classification under symmetric versus calibrated bands, all 41S11 retrieval ablations, best cost-MAPE configurationS12 five-fold cross-validation, best cost-MAPE configurationFigure S1 run-to-run reproducibilityText S1 system and user prompts No personal or identifiable data. Project-level source data are confidential and described in the manuscript.