Nabi X
(Last update: 15 June 2025)
Nabi X is a modular educational AI system developed by BDSL with the support of Teaching Development and Language Enhancement Grant (TDLEG) in conjunction with the department of History at Hong Kong Baptist University. It builds on the following interconnected endeavors:
- Nabi, an earlier incarnation of Nabi X supported by the University of Hong Kong’s Teaching Development Grant (TDG), is a participatory AI chatbot for digital humanities pedagogy based on retrieval-augmented generation (RAG), chain-of-thought (CoT) reasoning, and knowledge bases co-created by students from course readings. Nabi teaches students how to assemble and calibrate their own compound AI system, providing them with personalised assistance in return. Nabi’s modular design limits the role of a large language model (LLM) to a speech engine, sourcing knowledge primarily from vector embeddings and knowledge graphs extracted from user-provided documents.
- AI Philosophers enables students to engage in virtual conversations with two major philosophers, Zhu Xi (1130–1200) and Voltaire (1694–1778), with plans to introduce additional figures. Thanks to the multilingual capabilities of LLMs, students can explore unfamiliar philosophical traditions in any major human language of their choice—for example, China’s Neo-Confucian tradition in English or German, or the French Enlightenment in Korean or Japanese. Our carefully constructed word embeddings help minimise hallucinations, and all conversations include citations to relevant sources.
- DeepPast: A set of generalizable AI workflows aimed at automating key aspects of historical research, such as data extraction, named entity recognition, and contextualised interpretation.
Core Components of Nabi X
A. Participatory Knowledge Curation
Student-Built Knowledge Bases: Students convert theoretical literature and archival materials into machine-readable forms (e.g., vector embeddings, knowledge graphs), deepening their grasp of course content while developing skills in structuring unstructured data. This hands-on process prepares them to become AI co-creators within their chosen areas of specialization. By actively contributing to the creation of curated datasets, students also gain experience in responsible data stewardship and documentation, reinforcing both technical competence and critical reflection.
B. Secure and Compliant AI Ecosystem
Copyrighted Document Uploads: Copyrighted book chapters and journal articles are uploaded to on-campus servers, processed once solely for educational purposes, and immediately discarded in accordance with the licensing terms.
Personal Data Handling: Students’ personal work (e.g., graded essays, project drafts, coursework submissions) is processed through encrypted channels. No copies are stored off-campus, and nothing is retained beyond assessment unless the student explicitly opts in. All handling of personal data fully complies with university policies and Hong Kong’s Personal Data (Privacy) Ordinance.
C. Hallucination Mitigation
Instructor-Guided Validation: Instructors can incorporate curated materials and model prompts to guide responses to common student questions about assigned readings, especially where disciplinary tacit knowledge or assumptions may not be made explicit in the text.
Adjustable Parameters: Students can experiment with model settings (e.g., temperature) to explore how variations affect AI responses. The goal is to help them to see AI as a collaborative tool for brainstorming and learning and reduce reliance on it for definitive answers.
Reducing Zero-Shot Prompts: Regular engagement with a controlled AI learning environment discourages last-minute use of general-purpose LLM chatbots and the submission of generic written work prone to hallucination.
D. Curriculum Integration
Support for Shy Students: Q&A channels and virtual tutorials help foster a more inclusive and equitable learning environment where students can explore ideas at their own pace, without fear of judgment.
E. Sustainable and Ethical AI Development
Local Model Fine-Tuning: Instructors and students adapt open-weight LLMs (e.g., Llama, Qwen, DeepSeek) to discipline-specific tasks.
Open-Source Contributions: Where licensing permits, knowledge bases and fine-tuned models are pushed to public repositories.
Nabi is currently under development at BDSL. The alpha version is scheduled to launch in September 2025.
