A website assistant that answers only from published pages
A working demonstration of a public admissions and curriculum
assistant, and a technical report on how it was built, what was measured,
and what equipment hosting one would require.
6 min 34 secLive demonstration + summaryNarration: edge-tts
Chapters
What this is
A chat assistant grounded in a college's own pages
The assistant sits in the corner of a college website and answers questions
about admissions, degree programs, curriculum, cost, and policies. Every
answer is drawn from the institution's published pages, cites the page it
came from, and links to it.
The demonstration institution is fictional and every figure on its site is
invented, but the pages are structured the way a real business school's pages
are. The assistant in the video is answering live — nothing is staged.
It touches no student data. Admission requirements,
curriculum tables, and tuition schedules are already published on the open
web. There is no education record in the pipeline, nothing to de-identify,
and no argument that the data must not leave the institution.
What the report covers
Including what it does not solve
How the system works, stage by stage, and why retrieval beats fine-tuning for this
A stale page that outranks current policy — and why filtering, not prompting, is the fix
A retrieval defect found in testing that produced a fluent, cited, wrong answer
Guardrails mapped to the 2025 OWASP Top 10 for LLM Applications, and the cost of over-blocking
Hosting options with equipment and order-of-magnitude figures
A plain list of the limitations, including what remains unverified