ESHS/HSS Annual Meeting

RAG in Practice: Closing the Usefulness Gap for Humanities Research

• In Event: The Impact of the Digital Revolution on the History of Science and its Social Implications-4

English Abstract

This talk presents lessons learned in building QuHiRa (Quantum History Research Assistant) and deploying it in the writing of a multi-authored survey volume growing out of the 2025 Heraeus Symposium "Revisiting the History of Quantum Physics." Technically, QuHiRa is based on the RAG—Retrieval-Augmented Generation—principle, in which relevant documents are retrieved from a curated corpus and used to ground the AI model's responses in verified sources. While building functional RAG systems has become technically straightforward, most implementations focus on benchmarks that overlook the critical gap between a system that functions and one valuable enough for experts to depend on for serious academic work.
In developing QuHiRa, this gap has been addressed through two parallel strategies: technical innovations and expert engagement approaches. Technically, QuHiRa employs a complex agentic RAG architecture, including a deep search mode that provides contextually rich, historiographically nuanced responses. Beyond major architectural decisions, however, QuHiRa also depends on minor features often missing from proof-of-concept implementations, such as document grounding that lets users jump directly to retrieved source passages, or indicators when the AI has taken over passages verbatim.Expert engagement measures include channels for suggesting corpus extensions, integration of workshop discussions, and a quality guardian program where individual experts monitor questions in their specialization and provide feedback on answer quality changes. This enables systematic quality tracking while keeping individual burden low and collective impact high. QuHiRa demonstrates that meaningful AI innovation in the humanities does not primarily lie in initial implementation, but in iterative optimization guided by expert feedback and scholarly practice. The talk offers a practical model for how AI tools can move from experimental novelty to genuinely dependable infrastructure for humanities research.

Author

• Jochen Büttner