HKU open-source system turns textbooks into tutors
Researchers at HKU released an open-source project that converts textbooks into personalized learning systems using multi-agent reasoning, promising an end-to-end pipeline for knowledge tracing and tailored content recommendation. The social posts highlight it as a practical foundation for experimenters who want to turn curriculum materials into interactive, adaptive lessons (x.com).
Most study software does one of two things: it stores your files like a cabinet, or it answers questions like a chatbot. DeepTutor, released by the University of Hong Kong Data Science Lab, tries to do a third thing: turn your own textbook or notes into a step-by-step lesson that changes as you go. (github.com) To do that, the system first builds a knowledge base from uploaded files, including Portable Document Format files, Markdown files, and plain text. That gives the model a set of course materials to pull from instead of guessing from general internet knowledge. (github.com) The basic trick is called retrieval-augmented generation, which is a long name for “look it up before you answer.” It works like an open-book exam: the model searches the uploaded material, finds the relevant passage, and then writes the explanation. (github.com) DeepTutor adds a multi-agent setup on top of that search layer. Instead of one model doing everything, it splits the job among specialized agents, the way a tutoring center might separate lesson planning, live help, and end-of-session feedback. (github.com; mintlify.com) In its guided learning mode, one agent scans a notebook and picks 3 to 5 knowledge points in a learning order. A second agent turns each point into an interactive HyperText Markup Language lesson page, a third answers questions during that lesson, and a fourth writes a mastery summary with follow-up suggestions. (mintlify.com) That is the part people mean when they say it can turn textbooks into tutors. The software is not just extracting text from a chapter; it is reorganizing the material into a sequence, presenting it as a lesson, and keeping the chat grounded in the specific concept you are on. (mintlify.com) The Hong Kong team pushed version 1.0.0 on April 4, 2026 and described it as an “agent-native” rewrite under the Apache 2.0 license. The repository also shows 13,500 stars and about 1,800 forks, which is unusually fast uptake for a niche education tool. (github.com) The release notes show how quickly the project is moving. Between April 4 and April 8, 2026, the team shipped three beta updates that changed software development kit support, parsing, Windows compatibility, and guided learning fixes. (github.com) What makes this more than a prettier reader is the full loop. The same system that ingests documents can also generate explanations, answer context-specific questions, and produce personalized next-step recommendations at the end of a session. (github.com; mintlify.com) That still does not mean it has solved education. The public material shows a strong open-source foundation and a practical workflow for experiments, but it does not by itself prove better test scores, lower dropout rates, or reliable knowledge tracing across real classrooms. (github.com; unwire.hk)
