Neuro-Symbolic AI Improves LEED Compliance Document Screening

The process of achieving LEED v4.1 BD+C certification is highly document-intensive, requiring manual review of hundreds of pages and the application of specific credit logic. This research investigates the potential of small, locally deployed language models (LLMs) to automate meaningful screening of LEED documentation, exploring how deterministic symbolic components can complement this process. A novel neuro-symbolic pipeline was developed, which aligns project PDFs to LEED credit sections, retrieves relevant evidence using keyword signatures, verifies compliance with a local 4-billion-parameter LLM, and applies a specialized numeric checker for quantitative thresholds. Experiments conducted on four university buildings, involving 484 PDFs and 153 credit-level decisions, revealed that the 4-billion-parameter Gemma3:4b model was the most effective text-only verifier, achieving 67.3% accuracy and outperforming a larger 8-billion-parameter model. The deterministic numeric checker proved crucial for correcting arithmetic errors, boosting accuracy for quantitative credits like EA-p2 from 50% to 100%. However, the full neuro-symbolic configuration achieved 61.6% overall accuracy, slightly trailing the best text-only baseline due to extraction failures and conservative behavior on qualitative categories. Interestingly, adding low-resolution drawing images consistently reduced accuracy, suggesting that multimodal inputs are not always beneficial. The study also found that prompt effectiveness varied based on documentation richness, with rubric prompts excelling in data-rich projects and chain-of-thought prompts performing better in documentation-lean scenarios.