Context-engineered question answering for multi-format construction safety documentation

Construction safety documentation is essential for accident prevention and risk management. However, existing systems often struggle with context-dependent question answering (QA) over sensitive, lengthy, and structurally diverse corpora. To improve queryability, we propose a context-engineered QA workflow that explicitly models how contextual information is acquired, organized, and used across different data modalities. Using multi-format crane-accident documents, we test three hypotheses: large language model (LLM)-synthesized supervision adapts embedding-based retrievers to crane-safety terminology without manual labels (H1); a modular retrieval-augmented generation (RAG) architecture, which transforms content into query-conditioned representations, improves analytical answer quality (H2); hybrid manual and automated assessments capture error patterns at different granularities (H3). On a crane-safety benchmark, retriever fine-tuning raises Recall@1 by 14.0% over the untuned baseline and 10.6% over a widely used proprietary embedding model. At the system level, integration with Table-Lens and staged generation achieves the highest faithfulness, significantly outperforming controlled reimplementations of two recent multimodal RAG protocols. While automated metrics capture broad patterns, manual review surfaces line-level errors and remains essential for safety-critical decisions.