Justin DeMayo

Panel discussion on how AI can be used in research, with examples of chatbots and machine learning in global HIV research and an overview of AI resources available at Yale and elsewhere

Constructing a rigorous search strategy is not a single task—it is a specialized workflow: decomposing a research question, selecting an evidence-based framework, mapping terms to controlled vocabulary, assembling Boolean logic, and validating protocol compliance. This work presents ARQA (Automated Research Query Assistant), a multi-agent AI assistant currently in evaluation phase at the Harvey Cushing/John Hay Whitney Medical Library. Designed to augment biomedical research workflows, ARQA is built on the premise that distinct steps of the workflow can be best supported by specialized agents working in coordination rather than a single general-purpose model. ARQA's LangGraph pipeline assigns distinct responsibilities to four agents: a Reasoning Agent analyzes research questions using 38 frameworks (e.g., PICO, SPIDER) to assemble Boolean structures; a Semantic Search Agent resolves concepts to MeSH vocabulary using vector embeddings; a Protocol Agent ensures compliance with standards like PRISMA-P; and a Critic Agent independently validates the query, catching common Boolean logic errors—such as incorrectly combining distinct conceptual frameworks with OR—which severely degrade search precision. This architecture argues that multi-agent systems can encode the sequential, role-differentiated structure of search workflows in ways monolithic LLMs cannot. Grounded in natural language processing and health sciences librarianship, ARQA offers a case study in designing AI to effectively augment the work of a domain expert.