The University Knowledge Problem
How might we create a reliable system that allows students to find university information without searching across multiple platforms?
The problem
Where students run into it today.
Everything is written down somewhere. Nobody knows where.
Why it matters
The information usually exists. The cost is the search — a student ends up asking three people before finding a document that was published a month ago.
Answers that get passed around by word of mouth drift. A date remembered wrongly by one person becomes the date an entire group believes.
This is the clearest example on campus of a retrieval problem, which makes it an unusually good way to learn how retrieval systems actually behave outside a tutorial.
Possible directions
Several credible approaches, no obvious winner. Choosing between them is part of the work.
Retrieval over a curated corpus
Collect a bounded set of documents, chunk them carefully, embed them and answer only from what was retrieved. The interesting engineering is in chunking and in refusing to answer when confidence is low.
Structured extraction first
Rather than answering free text, pull dates, fees and deadlines into a structured store and let the interface query it. Less impressive in a demo, far more reliable in practice.
Citation-first interface
Treat the source document as the answer and the generated text as a pointer to it. Every response links back to the page it came from.
Freshness and expiry
A notice from last semester is worse than no notice. Model document validity explicitly instead of hoping the ranking handles it.
Technologies & difficulty
The approach is not known in advance. Expect a literature pass, prototypes that get thrown away, and results you have to argue for.
Difficulty describes the work, not the person. A level 04 problem is not closed to a first-year — it just needs a team and a longer runway.
Potential project
Campus Knowledge Assistant
A question-answering layer over a curated, permission-aware corpus of university documents that always cites its source and declines when it does not know.
Skills needed
Open roles & possible team
- AI / LLM
Retrieval pipeline, chunking strategy, evaluation harness
- Backend
Ingestion, storage, document freshness and access rules
- Frontend
Query interface, citation rendering, empty and refusal states
- Research
Answer quality methodology and failure analysis
Research questions
The things nobody in the group can currently answer.
What chunking strategy survives documents that mix tables, dates and prose?
How do you measure whether an answer is correct when there is no labelled dataset?
Can the system be made to say “I don’t know” often enough to be trusted, without becoming useless?
What is the right unit of freshness — the document, the section, or the individual fact?
Next steps
What the first two weeks would sensibly look like.
- 1
Assemble a corpus of 50–100 publicly available documents and label 40 real student questions against them.
- 2
Build the smallest possible retrieval baseline — keyword search, no model — and record its accuracy.
- 3
Only then introduce embeddings, and justify the improvement with numbers.
- 4
Write up what broke. The failure list is the contribution.