Deploying Conversational Agents in Virtual Research Environments: Approaches and Lessons Learned

The rapid progress of conversational artificial intelligence and Large Language Models (LLMs) has opened new opportunities to enhance user interaction, support, and accessibility in Virtual Research Environments (VREs). This poster presents the approaches, challenges, and lessons learned from a multi-year e!ort to design, develop, and deploy conversational agents within the D4Science infrastructure. Through three successive implementation cycles—Janet, D4Science AI Agent, and DAVE—the poster traces a process of iterative refinement aimed at improving flexibility, extensibility, usability, and integration with existing VRE services. Janet, the first prototype, explored modular NLP components but revealed limitations in adaptability and feedback integration. The second approach, based on the Cheshire Cat framework, improved modularity and LLM interoperability but remained constrained by a single-agent design. The latest solution, DAVE (D4Science Assistant for Virtual research Environments), introduces a multi-agent architecture built with Google’s Agent Development Kit, enabling secure and context-aware interaction with multiple D4Science services. DAVE combines specialized agents for tasks such as document analysis, catalogue navigation, social interaction summarization, and algorithm deployment within D4Science’s computational platform. Integrated feedback mechanisms and a Retrieval-Augmented Generation (RAG) knowledge base further enhance its learning and personalization capabilities. The findings demonstrate that conversational agents can lower barriers to VRE adoption, streamline workflows, and foster user engagement by o!ering intuitive, natural language interfaces. Lessons learned from this evolution suggest key design principles for future research infrastructure agents, emphasizing modularity, interoperability, and data security. Future work will involve usability evaluations, the integration of user-driven feedback, and experimentation with locally-hosted LLMs to strengthen privacy and operational sustainability.