The intelligent management of hazardous chemicals inherently relies on high-fidelity, structured safety knowledge. However, the vast majority of this critical information remains trapped in highly unstructured, complex regulatory documents, notably Material Safety Data Sheets. Extracting actionable knowledge from these fragmented tables and verbose technical texts poses a formidable challenge. Traditional rule-based natural language processing pipelines are excessively brittle, whereas employing monolithic large language models invariably introduces fatal factual hallucinations in zero-tolerance safety scenarios. To overcome these critical bottlenecks, we propose the Multi-Agent Knowledge Extraction and Reasoning (MAKER) framework, an automated, industrial-grade solution. MAKER systematically decomposes the complex extraction pipeline into granular, specialized tasks through a stateful multi-agent paradigm. It seamlessly integrates the deterministic precision of regular expressions with the contextual reasoning of a specialized agent society-comprising extractors, normalizers, critics, and linkers. Through a continuous evaluate-feedback-refine loop, the system eliminates error cascading and ensures strict schema compliance. Using this interactive framework, we constructed one of the most comprehensive multimodal knowledge graphs for hazardous chemicals to date, encompassing 2,380 core entities, over 31,000 nodes, and 108,000 relations. To definitively demonstrate its industrial value, we integrated this highdensity graph as a structural prior into KGC-CLIP, a vision-language pre-training model for emergency chemical identification. Rigorous downstream evaluations in a critical visual retrieval task reveal that without the MAKER-constructed graph, the visual model completely fails due to visual ambiguity, yielding a near-zero top-1 recall of 0.22%. Conversely, incorporating the graph propels the model's performance to a substantial accuracy of 62.86%. These results conclusively demonstrate that MAKER provides the indispensable spatial skeleton required for robust multimodal AI applications in high-stakes scientific domains.