ChatHA – From an Information Retrieval Agent to a System of Expert Agents

Abstract

Scientific research increasingly relies on heterogeneous information sources, creating a strong motivation for individualized, context-specific chatbots that support researchers in navigating domain-specific knowledge. However, a central challenge is that researchers sometimes lack a clear understanding of what questions to ask, particularly when exploring unfamiliar topics at the beginning phase of research or unknown datasets, and at the same time require transparency and traceability of the system’s responses.<br/><br/>To address these challenges, ChatHA (Humanities-Aligned Chatbot), a Large Language Model (LLM) based chatbot, was developed that integrates with some research data repository (RDR), like Zenodo or the local RDR at the University of Hamburg, to present information not only present in the metadata but also in the data of the entries themselves. Researchers may now be interested to also publish their own version of ChatHA based on their own research work. A user potentially wants to integrate and discuss the information provided by multiple instances of ChatHA simultaneously, e.g. in an interdisciplinary environment, where information from multiple fields and other researchers is important. In this paper, we introduce a new iteration of ChatHA that functions as a system combining multiple versions of ChatHAs generated and shared on RDRs by other researchers. This is achieved by basing this umbrella ChatHA on the methods from Co-STORM (collaborative Synthesis of Topic Outlines through Retrieval and Multiperspective Question Asking). Co-STORM employs multiple LLMs agents, each of which taking the role of an expert in a specific field. A special LLM agent takes the role of the moderator to create a situation similar to that of a roundtable discussion, where the moderator is leading the discussion into specific directions and is choosing which expert agent is next to provide input. The user of the system can see all the discussions and can intercept at any step to ask follow-up questions or clarify aspects.<br/><br/>The approach is illustrated using a use case in which ChatHA is applied to a predefined set of research-related web pages, here artefact profiling guide, enabling dialogue-oriented search and structured exploration of these sources. The results show that the combined ChatHA and Co-STORM approach improves information discovery and research efficiency by supporting transparent search and helping researchers formulate relevant questions.

Ähnliche Arbeiten