"To provide nursing educators and students with accurate and efficient AI-powered support, the EBN-Expert system was developed using Coze, a low-code platform that enables the creation of intelligent conversational agents through graphical interfaces and modular components, requiring minimal manual programming. ... This design supports ...metacognitive processes such as structured reflection [via] built-in integration tools [that] connect to: OpenAI’s text-embedding-ada-002 model for vectorization; Pinecone for semantic search; Bing and Baidu application programming interfaces (APIs) for web retrieval; Deepseek-R1 for language generation. ... The overall reasoning workflow, illustrating how these components operate in sequence, is visualized in Fig. 1."

Fig. 1. Workflow design of the evidence-based nursing expert (EBN‑Expert). "The Reflection Module operates within a single-agent architecture and is activated immediately following the generation of a candidate answer. Its primary function is to evaluate both the answer and the associate reasoning trace.... The module uses three knowledge sources: an internal embedded textbook corpus and two external web-based search tools—Bing and Baidu—via APIs." "The Decision-Making Module... operates using a structured hard voting mechanism. It receives three independent reflection outputs—each comprising an Output field ('Sufficient' or 'Insufficient') and an associated reasoning trace—as produced by the Reflection Module (see Box 3). If most outputs (i.e., at least two out of three) are labeled “Sufficient”, the candidate answer is classified as 'Correct'. ... If the decision is “Incorrect”, the system re-initiates the retrieval–reflection–decision cycle... enabling iterative refinement...."

"As shown in Table 1, the overall performance scores differed significantly among the four models (P < 0.001). The EBN-Expert achieved the highest mean score of 1035.00 (SD 17.32), with a minimum of 1025 and a maximum of 1055."

"The results of multiple comparisons, as presented in Table 2, demonstrate that the EBN-Expert significantly outperformed DeepSeek, Kimi and ChatGPT-o1 (P < 0.05 for all comparisons). Additionally, DeepSeek achieved a significantly higher score than Kimi (P < 0.001), while ChatGPT-o1 also outperformed Kimi (P = 0.009). However, the difference between DeepSeek and ChatGPT-o1 was not statistically significant (P = 0.084)."