ClinLab-RAG: A Multimodal RAG System for the Early Verification of Risks in Medical Equipment Supply Chains

Checking manually whether a reagent works on a given analyser takes a median of 12.3 min at the distributor we studied. We developed ClinLab-RAG, a retrieval-augmented generation architecture with hybrid search, a product compatibility graph and five LangGraph agents, and evaluated it through Design Science Research on 394 technical and ERP fragments and 175 queries. In the automated evaluation, the submitted architecture raised answer relevancy from 0.416 (keyword retrieval) to 0.522 while faithfulness fell from 0.945 to 0.724; component analyses attribute the gain to the specialised prompt and show that the graph raised context recall at the cost of faithfulness. A revised architecture checks compatibility against the manufacturer’s matrix and hazard classifications against the source safety data sheet deterministically. It raised context recall from 0.416 to 0.460, gave correct verdicts for every recognised equipment–reagent pair and included the complete GHS classification in 14 of 17 safety answers, against 1 of 17 before. Three blinded domain experts rated v2 and the strongest baseline near the ceiling (exactness 1.95 vs. 1.89 of 2; p = 0.180). The system supports verification before dispatch; its effect on procurement errors remains to be measured.