EthosPowerAI
Meet DeepTutor and its publisher
DeepTutor is an open-source tutoring application from the Data Intelligence Lab at the University of Hong Kong, published as HKUDS/DeepTutor under the Apache-2.0 licence, with a Python backend, a Next.js web interface and a site at deeptutor.info. The README names Bingxi Zhao as lead maintainer, with Xingyu Hou and Jiahao Zhang. On 7 November 2026 the repository had 55,949 stars, 6,845 forks and 316 open issues and pull requests; it was created on 28 December 2025, its last commit landed on 6 November 2026, and release v1.7.2 shipped the same day. Releases arrive every few days, and an April 2026 rewrite touched roughly 200,000 lines. No paid edition exists. It sits in my open-source catalogue.
List what it actually does
You upload PDFs, Word files, slide decks and spreadsheets into a knowledge base, and a chat tool called rag answers from them with citations you can trace. The retrieval engine is a choice: LlamaIndex hybrid search by default, PageIndex for page-level citations, GraphRAG or LightRAG, with MinerU as an optional parser, and a kb eval command scores retrieval against your own questions. On top of that sit a quiz generator, an exam mimic that imitates a reference paper, a mastery path that breaks a goal into topics with a hard gate on each, spaced review of the learner's mistakes, and a 'living book' compiled from the sources. It runs against local models through Ollama, LM Studio, llama.cpp or vLLM, or against the usual cloud providers.
Correct what it is for
HKUDS built DeepTutor for students, and the gaps show the moment a plant tries to train staff with it. Authentication is off by default; the optional multi-user mode keeps local usernames and passwords, makes the first registered user the administrator, and has no single sign-on, since a search of the source on 7 November found no Entra, OIDC or SAML code. It has no SCORM, no xAPI, no certificates and no competency records. Its audit file, an append-only log of resource access and admin actions, records who opened what; it does not record who is competent to start up the line. A log is not a register. Microsoft 365 support amounts to one Teams chat channel, with no SharePoint connector.
The headline numbers need the same care. The team's paper, arXiv 2604.26962, reports a 10.8% average gain on personalised metrics and 29.4% on general agentic reasoning across five model backbones, measured on the authors' own TutorBench with an LLM playing the student, over university curricula. Those figures describe the authors' benchmark. They say nothing yet about a shift operator learning a line start-up.
Design a tutor inside the fence
The solution puts DeepTutor to work for a battery-materials pilot plant of about thirty people making graphite anode material for lithium-ion cells, scaling towards continuous production on Microsoft 365 with one IT person, and protective of patented process know-how. That last fact sets the architecture: DeepTutor in Docker on a server on site, local models through Ollama, web search and every outbound feature switched off, and the host firewalled. The project has a DuckDuckGo search fallback in its code and does not document its default traffic, so the design confirms with a packet capture during commissioning that nothing leaves.
The document controller approves which revision of each SOP, work instruction, safety data sheet and maintenance manual is current in SharePoint, and a scheduled sync copies only those revisions into DeepTutor's linked folder and removes anything superseded, so that the operator who asks at three in the morning how to purge a line before a restart gets the procedure the plant approved last week, not the revision it withdrew last month that may still sit in somebody's downloads folder. PageIndex makes each answer point at a page, and supervisors write the questions kb eval must pass before anyone learns from the knowledge base. Operators, maintenance technicians and lab technicians each get a mastery path for their role. When DeepTutor's gate passes, a supervisor still watches the task done on the real equipment and records the sign-off in a SharePoint list, the training register the plant already trusts. Questions the tutor could not ground go to the coordinator as gaps in the documents. I would pin the version and upgrade quarterly, after kb eval runs again.
See the design, its diagram and the demo at the solution page.
The tutor can teach the SOP. A supervisor still signs the operator off.