Retrieval Assistant for University Regulations
A question answering assistant for students and staff over a university's own handbook, examination rules and fee policies, answering in plain language and citing the exact clause it used. A RAG final year project where the interesting engineering is refusing to answer when the documents do not cover the question.
Intermediate
A small team, or one strong student willing to learn something new.1 semester · 2 students
7 to ship
4 optional extrasSuggested stack
What you should ship
- An ingestion pipeline that splits the official PDFs by section and clause, storing page, section number and document version with each passage
- Hybrid retrieval and an answer service that returns an answer plus the exact passages it was drawn from
- A web interface showing each citation linked to the source page, with the quoted clause visible
- A test set of at least 100 real student questions with expected answers and supporting clauses, plus 20 questions the documents do not answer
- Retrieval accuracy, answer accuracy and refusal accuracy reported separately
- Refusal behaviour: out-of-scope questions get a clear "not covered" reply that points to the right office
- An admin screen to upload a new version of a document and see which test answers changed
If you have time left
- Versioned passages so an answer can be traced to the policy in force on a given date
- Urdu questions answered from English documents, evaluated on a translated subset of the test set
- A confidence score shown to the user, calibrated against the test set
- A WhatsApp or Telegram front end for the same service
The problem students have with university rules
Every university has a handbook nobody reads, a regulations document nobody can navigate, and an office that answers the same questions all year: How many times can I repeat a course? What is the fee refund rule if I withdraw in week three? Can I take a summer course to improve my CGPA? The information is public and correct; it is simply hard to find, spread across PDFs, notices and scanned circulars.
Wrong answers from seniors or WhatsApp groups have real costs: missed deadlines, lost fees, ineligibility for exams. Staff time goes to repeating answers instead of handling exceptions.
What already exists and the gap
- Mainstay and Ivy.ai sell chatbots to universities, mostly in the United States, for admissions and student services.
- Gemini Notebook (formerly NotebookLM) lets anyone upload documents and ask cited questions about them.
- General chat assistants answer university questions confidently, often from outdated or unrelated sources.
Commercial products are priced and designed for large institutions with budgets. General tools will answer anything, including what the handbook does not say. The gap is a small, institution-owned assistant that cites exact clauses, knows which version of a policy it is using, and reliably says "not covered, contact the examinations office" when the documents are silent.
What you build
Users: students asking questions, and an admin in the registrar or examinations office who uploads documents.
- Ingestion: parse PDFs, split by clause rather than fixed-size chunks, keep section numbers and page references.
- Retrieval: hybrid keyword and embedding search, since clause numbers and course codes match better with keywords.
- Answering: a prompt that must answer only from retrieved passages and cite them, or refuse.
- Admin: upload a new version, rerun the test set, see what changed.
The technically interesting part
The hard problem is refusal. Any system like this will confidently answer questions the documents do not cover, and that is the failure that matters, because a wrong answer about exam eligibility is worse than no answer. Combine a retrieval score threshold, an explicit "is this answered by these passages?" check, and a test set of deliberately unanswerable questions to tune both. Show your evaluators examples of correct refusals and the few that slipped through.
Clause-aware chunking is the second problem. Regulations cross-reference each other ("subject to clause 7.3"), and splitting mid-clause loses the conditions that change the answer.
How to evaluate it
Build the question set early, before the system.
- Retrieval accuracy: is the correct clause in the top 5?
- Answer accuracy on the 100 answerable questions, graded by two people.
- Refusal accuracy: share of the 20 unanswerable questions correctly declined, and share of answerable ones wrongly declined.
- Citation validity: every cited passage actually supports the answer.
- User test: 10 students ask their own questions; record whether answers were correct and whether they would trust it.
Suggested plan
- Weeks 1 to 2: collect documents and 120 questions from students and staff, with answers and clauses.
- Weeks 3 to 5: ingestion and clause-aware chunking; retrieval with measured accuracy.
- Weeks 6 to 8: answering with citations; refusal logic.
- Weeks 9 to 10: interface and admin screen.
- Weeks 11 to 13: full evaluation and user test.
- Week 14: report and demo.
Path to a product
The first customer is your own university: pitch it to the registrar or student affairs office with the evaluation results. Other universities, colleges and schools have the same problem, so offer it as a hosted service per institution with an annual fee, including document ingestion and a refreshed test set each semester. The same design fits HR policy handbooks at companies.
Scope warning
- Do not train or fine-tune a model. The project is retrieval, grounding and evaluation.
- Do not connect to student records or personal data.
- Do not answer beyond the documents, even when a question is popular.
- Get permission from the university before publishing their documents in a public demo.
Learn what you need on Sefism
- AI engineering roadmap : RAG, embeddings and evaluation in order.
- AI engineering cheat sheet : chunking, retrieval and prompting patterns at a glance.
- PostgreSQL Documentation : full-text search and pgvector in one database.
- Gemini Notebook : try cited question answering over your documents before building your own.
- Which AI topics are worth learning : why RAG done well is a hireable skill.
Get help with this idea
The strongest version of this project is built around its test set and its refusals. Tauseef can help you plan the question collection, the chunking approach and an evaluation your university would trust enough to adopt the tool. If you are new here, create a free account first.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.