Figure 2
This figure presents a layered conceptual architecture of an Agentic Publication system, organized into three interconnected levels that together support adaptive, query-driven access to scientific knowledge. The bottom layer represents research and knowledge generation, including raw and processed experimental data, simulations, analytical workflows, metadata, and expert interpretations. These elements are structured and stored in a knowledge base designed to preserve provenance, context, and updateability rather than being reduced to static text. The middle layer consists of an L L M-driven processing and retrieval-augmented generation (R A G) system. This layer integrates large language models with structured retrieval, ensuring that generated responses are grounded in verified knowledge rather than unsupported inference. It manages semantic interpretation of queries, retrieval of relevant knowledge units, synthesis of responses, and consistency across evolving content. Feedback loops illustrate how new results, corrections, or reinterpretations can be validated and reintegrated into the knowledge base, enabling continuous refinement. The top layer represents user interaction, encompassing both human users and machine agents. Users may submit queries at different levels of abstraction, from conceptual overviews to highly specific technical questions. The system dynamically adjusts the granularity of responses, returning summaries, explanations, references, or links to underlying datasets as appropriate. This layered architecture highlights how Agentic Publications enable flexible semantic resolution while maintaining transparency, traceability, and long-term maintainability.Conceptual illustration of the proposed Agentic Publication (inspired by the “PLOS-LLM” model by Hughes and Van Heerden, 2024). New research and knowledge generation steps feed data into the LLM-centric system, while user queries retrieve synthesized answers. The system allows users to zoom in and out on the level of detail – from high-level summaries (headlines/abstracts) to granular data (complete datasets). An interactive loop based on interaction log analysis helps keep knowledge updated and easily accessible