The model stops being the source of truth and becomes the reader. That is the entire idea.
A retrieval system finds the passages in your documents relevant to the question, and a generation system answers using only those passages, citing them. That is what makes AI usable on contracts, policies, regulations and client documentation where invention is unacceptable. Tools today: NotebookLM, Expertex.ai, Semrush, Opal.
Mechanically, documents are cut into chunks, each chunk becomes a vector, a numerical representation of its meaning, and your question becomes one too. The system retrieves the chunks closest in meaning, not in keyword, which is why it finds the right passage even when your wording differs entirely from the document’s.
Chunking matters: cut too small and a clause loses the context that gives it meaning, cut too large and retrieval gets imprecise. Retrieval failure is invisible: if the right chunk is not retrieved, the model answers confidently from the wrong ones. A citation proves where a claim came from, never that the right source was found.
A grounded assistant must decline when the documents do not contain the answer. Test this deliberately: ask something your corpus genuinely does not cover, and if it invents, tighten the instruction until it refuses cleanly. Then decide the second refusal class: what it must not answer even when the documents contain the answer. Salaries, disciplinary records, legal positions, anything commercially sensitive. That is a policy decision, and it belongs in writing before deployment.
Plant a wrong fact in a source document and your assistant will repeat it, with a citation, forever. Grounded systems inherit the honesty of their corpus exactly. So name an owner: someone responsible for what is in the knowledge base, for removing superseded policies, for versioning, for the update process when a regulation changes. If nobody owns it, the system is accurate on launch day and quietly wrong six months later, which is worse than no system, because people have started trusting it.
Anything running unattended needs a way to be seen. Decide up front what is measured, usage, refusal rate, unanswered questions, response time, cost, what the threshold is and what you do when it is crossed, and who looks, how often, in a calendar, not an intention. The unanswered questions log is the most valuable artefact, it tells you precisely what your knowledge base is missing, in the users’ own words.
The final deliverable is a business case, not a demo, because that is the actual job. Cost, risk, the human in the loop policy from Day 13, the update process, the monitoring plan, and an honest statement of what the system will not do. Anyone can show a working assistant. Very few can show one with an owner, a refusal policy, a monitoring plan and a maintenance process attached. That gap is where the commercial work is in this market.
NOVA reacts, nothing is scored, nothing is stored against you.
Load your subject notes and set texts into NotebookLM and interrogate them. Ask questions the documents do not answer and confirm it declines instead of inventing. That refusal is the whole point of the technique.
Build the capstone. A grounded knowledge assistant over a real document set, with source citations, a defined refusal behaviour, and a documented update process. Wrap it in the Day 12 automation so the knowledge base refreshes on a schedule. Put monitoring on the public output. Then present it as a fifteen minute business case with cost, risk and the Day 13 human in the loop policy.
Day 15 in progress
That is the fifteen days. Next stop: the final examination. NOVA will be in the room, quietly.