Student services assistant
An assistant that answers policy and administration questions from a university's own regulations and news, and cites the page each answer came from. Administrators keep the sources current, and a reviewed set of test questions is run before any change ships.
- Started
- May 2026
- Development time
- 7 weeks
- Platforms
- Web
- Industry
- Higher education
- −38%
of the routine questions that used to reach senior students and the office
- 91%
of questions answered from the regulations themselves, with the page cited
- 232
official documents behind the answers, kept current by the office
- 4min
for an administrator to publish an updated regulation, without a developer
The challenge
Questions about credits, deadlines, fees, and exams are answered somewhere in hundreds of pages of regulations, and a generated answer is only useful if the reader can check it. The documents change every semester.
The solution
Students ask in plain language and get an answer with the document and the page it came from, so every claim can be checked against the source. Administrators keep the sources current without a developer.
- Cited answersEach answer links the document and the page it came from, with a confidence level.
- Follow-up questionsA follow-up is rewritten into a standalone query using the chat history.
- Source managementAdministrators sync document links, upload PDFs, and rebuild the index.
- Index snapshotsEach reindex creates a new snapshot, so a rebuild can be rolled back in seconds.
- Local or hosted modelsEmbeddings and generation run on the university's servers or through a hosted provider.
Engineering decisions
Python
FastAPI
LangChain
React
TypeScript
Docker
- Hybrid retrieval: vector search (FAISS) fused with keyword search (BM25) and capped at three passages per document, then reranked by a cross-encoder. Each alone missed either exact names or paraphrases.
- The model checks its answer against the retrieved passages before replying, and states a confidence level next to the citations.
- 77 reviewed questions and a model judge with confidence intervals run against every change. At 0.64 grounded correctness against a gate of 0.8, the assistant is presented as a reader's aid, with citations to check.
- Docling parses the PDFs with page metadata, so citations survive a reindex, and the whole pipeline runs from one compose file on the university's servers.