How the AI Works

Short version: it is not allowed to answer from memory.

A general-purpose chatbot dropped into a student group will answer anything, confidently, including things outside your syllabus and things that are simply wrong. That is not a prompt problem — it is what a model does when nothing constrains it. Our pipeline constrains it at every stage.

The pipeline

Route. The question is classified — which subject, which level, is it even a question. Anything that is not gets handled without spending a model call on it.
Retrieve. A hybrid search over the institution's own material: dense vector search for meaning, sparse keyword search for exact terms like a section number, then a reranking pass to put the genuinely relevant passages on top. Two methods because either alone misses a different category of question.
Assemble. The retrieved passages are ordered by authority — the institution's own notes outrank general reference material — and become the only context the model is given.
Answer. The model writes an answer from that context. It is not asked what it knows; it is asked what the sources say.
Verify. A second model checks the drafted answer against the same sources. A claim not supported by them is sent back.
Repair. The answer is corrected and re-checked before the student sees it.

The guardrail

Every incoming message passes a guardrail before it reaches the model, built against the OWASP LLM Top 10 — prompt injection and prompt leaking in particular. It carries a canary token to detect system-prompt extraction, refuses in the language it was addressed in, and fails closed: if the guardrail itself errors, the message is refused rather than passed through.

Humans in the loop

Dedicated monitors review answers as they go out. A student who is not satisfied can say so, and the question is routed to a human expert rather than retried on the model. The AI reduces the volume that needs a person; it does not pretend to replace one.

Where it runs

Fully managed by us — servers, models, retrieval index, updates. The institution supplies material and gets a working system; there is nothing to deploy and no API key to hold. All model and infrastructure credentials stay on our side and are never embedded in any client application.

What we build → · Talk to us →