Adaptive hybrid AI-based retrieval-augmented generation for construction and seismic regulatory question answering

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Construction and seismic regulatory documents, such as the Uzbek QMQ (Qurilish Me’yorlari va Qoidalari – Construction Norms and Rules) and ShNQ (Shaharsozlik Normalari va Qoidalari – Urban Planning Norms and Rules), contain structured clause identifiers, numerical constraints and domain-specific terminology that conventional dense retrieval methods often fail to match reliably, while retrieval-augmented generation (RAG) systems additionally suffer from retrieval latency that limits interactive use. This paper proposes an Adaptive Hybrid RAG Model (AHRM) that combines semantic (Sentence-BERT with FAISS), keyword-based (BM25) and rule-based retrieval within a non-linear fusion function whose weights are assigned adaptively according to the detected query type, and deploys it as a real-time question-answering chatbot for Uzbek construction and seismic regulations. On a corpus of 120 regulatory documents (approximately 85,000 text chunks), AHRM reaches a Precision@3 of 0.88, an MRR of 0.85 and an answer accuracy of 91 %, outperforming dense-only, keyword-only and statically weighted hybrid baselines, with a mean index-search latency of 0.6 ms per query over the filtered candidate pool, measured separately from query embedding and answer generation. The results indicate that combining rule-aware retrieval with query-adaptive weighting improves the reliability of regulatory question answering in safety-critical engineering domains.