AI tutors grounded in course material: connecting course files to an assistant

What it means for an AI tutor to be grounded in course material

A grounded AI tutor is an assistant whose answers are built from a specific body of material, here the course's own files, instead of whatever the language model learned during training. When a student asks about a topic, the tutor looks in the lecture slides and notes for that course and explains what they say.

The difference matters because a general chatbot is fluent about almost everything and accountable for nothing. It may define a term differently from the lecturer, use notation the course never uses, or explain a method the exam does not test. A student cannot easily tell when that has happened. A tutor that works from the course file, and shows which file it used, keeps the explanation tied to what the student will actually be assessed on.

In Oxygen Academy, the university-aligned learning platform FLUIDS designed and built for students in Saudi Arabia, this is the organizing idea. The course team uploads and organizes course files, and tutors are associated with selected files so that different assistants can focus on different parts of the material. The tutor lives inside the course, next to the files, exams and summaries built from the same source.

How a course-grounded tutor works, step by step

Most grounded tutors use retrieval-augmented generation: the system finds the relevant passages first, then asks the model to answer using them. The same pattern underpins sourced religious assistants, as described in how an AI tafsir app should work; only the source material changes.

1. The course team uploads files. Slides, PDFs and notes are stored with the course they belong to.
2. Text is extracted and split. Each file is converted to text and divided into passages small enough to retrieve precisely, keeping the file name and the page or slide number with each passage.
3. Passages are indexed. Each passage is stored with a search representation, often a vector embedding, so it can be found by meaning as well as by keywords.
4. The student asks a question. The system searches only the files assigned to that tutor and picks the closest passages.
5. The model answers from those passages. Its instructions limit it to the retrieved material, ask it to cite the file, and tell it to say so when the material does not answer the question.
6. The answer is shown with its sources. The student sees which file, and ideally which page or slide, the explanation came from.

How to scope a tutor: one per course, per unit or per file set

The scope of a tutor is the set of files it is allowed to search, and choosing it is a teaching decision as much as a technical one. A narrower scope gives more focused answers; a wider scope answers more questions but retrieves less precisely.

Oxygen Academy associates tutors with selected course files, which lets the course team decide what each assistant should focus on. For a similar platform, show students the scope plainly, for example by listing the files a tutor works from, so they know which assistant to ask and why one may not know about a topic.

When a file is replaced or removed, re-index it and check which tutors depended on it. An index that still contains last semester's slides is a quiet source of wrong answers.

Citations and honest gaps: what the student should see

Every tutor answer should tell the student where it came from. A citation turns an answer the student has to trust into one they can check, and it sends them back to the material their exam is based on.

• File name and location. The file title and, where extraction allows, the page or slide number, linked to the course material viewer.
• A short quoted passage. For definitions and formulas, showing the exact source line avoids a paraphrase that subtly changes meaning.
• A clear 'not in your material' answer. When retrieval finds nothing relevant, the tutor should say the course files do not cover it and suggest asking the instructor, rather than answering from general knowledge without saying so.
• A label on general explanations. If the product allows the tutor to add background beyond the files, mark that part as general explanation, not course content.
• A report button. Students should be able to flag an answer as wrong, with the question, answer and sources attached for the course team.

Bilingual courses add one more rule. When the files are in English and the student asks in Arabic, the tutor can explain in Arabic, but it should keep the English term from the slides alongside the Arabic explanation, because that is the term that will appear in the exam.

Where course-grounded tutors fall short, and how to review them

Grounding narrows the ways a tutor can go wrong, but it does not remove them. Course AI should be presented as a study aid, and the platform should be honest about its limits.

• Extraction gaps. Scanned pages, equations, diagrams and tables often extract badly. If the text is wrong, the tutor's answer will be too.
• Wrong or blended answers. A model can misread a passage, combine two passages that do not belong together, or add a confident sentence the file does not contain.
• Outdated material. If files change mid-semester and the index does not, the tutor teaches the old version.
• Academic integrity. A tutor that solves graded assignments step by step can undermine assessment. Decide with the course team which kinds of help are allowed.
• Over-reliance. An explanation is not the same as understanding. Practice through exams and flashcards gives students a way to test themselves.

Review works best as a routine, not a one-off demo. Ask the course team to write a fixed set of real student questions per unit, record the file and page each answer should come from, and grade the tutor's answers for accuracy, correct citation and appropriate 'not covered' responses. Re-run the set when files, prompts or the model change. The same discipline applies to religious assistants; see AI guardrails for Islamic apps.

Planning an AI tutor for your platform

A tutor is easier to build well when it is one part of a course product rather than a standalone chatbot. The course already defines who may see which files, who the instructors are and what students are studying, and the tutor should reuse those rules instead of inventing its own.

• Check course access before the tutor searches anything, using the same rule as the materials page.
• Keep chat history per student and per course, and treat it as personal data under Saudi Arabia's Personal Data Protection Law (PDPL).
• Set usage limits per student so a busy exam week does not produce surprise AI costs.
• Give the course team a simple view of reported answers and of which files each tutor uses.

The tutor also works best next to other study formats built from the same files: exams with review and feedback, as in AI exam generation with a review step, and published summaries. The Oxygen Academy case study shows these pieces in one course experience, and our page on AI features inside your product explains how FLUIDS scopes AI work. Scope, price and timeline are agreed in a written proposal before development.