Figure 4
This figure shows a screenshot of the Voiceflow development environment implementing an Agentic Publication workflow based on retrieval-augmented generation. The visual layout depicts a structured sequence of processing blocks that govern how user queries are handled. Incoming queries are first analyzed and optimized to clarify intent, scope, and required level of detail, reducing ambiguity and improving retrieval accuracy. The optimized query is then used to search a curated knowledge base associated with the publication. Instead of retrieving entire documents, the system identifies relevant knowledge units such as data summaries, methodological descriptions, or validated interpretations. These elements are passed to a large language model that synthesizes a coherent response grounded explicitly in retrieved content. The workflow includes conditional logic to handle cases where relevant information is incomplete or unavailable. In such situations, the system generates structured fallback responses that transparently communicate limitations, suggest related queries, or indicate where evidence is missing. This behavior is explicitly encoded rather than emergent, demonstrating controlled and auditable system behavior. By visualizing this workflow, the figure illustrates that an Agentic Publication is an engineered system with explicit decision paths, governance, and error handling. The modular structure supports reproducibility, extensibility, and scientific accountability, reinforcing the practical feasibility of agentic publishing using existing orchestration tools.The Voiceflow development environment, a workflow to respond to the reader's queries using RAG. A high-resolution version is available at: https://doi.org/10.34965/I60500