A student asks a question and something confident comes back. Fluent sentences, a patient tone, an explanation that sounds right. None of that tells the student — or the teacher behind them — whether it is true. An answer with no source is a claim. The difference between a tool a teacher can allow in class and one they cannot is whether that claim can be checked against the book already open on the desk.
What a citation looks like here
Under a substantive answer, the platform names where the material came from: subject, book, chapter and page — something in the shape of "Mathematics — Grade 9 Maths, Ch. 4, p. 61". Where the page printed in the book differs from the scanned page, the printed one is shown, because that is the number a student will actually find. The citation is not decoration. It is an instruction to go and look.
A student sees one source: the one from their own grade. Three lines of provenance is noise to a child, and pointing one at a chapter two grades below is worse than citing nothing. Staff see the full list, because for them the spread is the signal — it shows what was actually matched, which is what makes a bad citation reportable rather than invisible.
Looked up when you ask, not baked into a model
This distinction is worth being precise about: it is easy to get wrong, and it changes what the system can honestly promise. The CDC and NEB textbooks were not used to train a model. There is no custom, proprietary or fine-tuned model here. What exists is a corpus — the books scanned, read and split into passages, with the grade, subject, chapter and page kept alongside each one.
When a question arrives, that corpus is searched — at that moment, for that question — and the passages that come back are what the answer is built from. Nothing about the curriculum is memorised into model weights. A page that turns out to be wrong can be corrected, and the next answer reflects it immediately. A book replaced by a new edition is re-ingested; nothing is retrained. And because the passage is retrieved rather than recalled, there is always a specific piece of text to point at — which is what makes a citation possible at all.
Two different questions, two different bars
Retrieval always returns something. The useful work is judging what it is worth, and the platform asks two separate questions about it rather than one.
- Is this question inside the syllabus at all? A looser threshold — its failure mode is refusing a legitimate question.
- Is the best match strong enough to quote and cite? A stricter one — its failure mode is a fabricated citation, an answer that says "the book says" when the book does not.
Falling short of the second bar does not mean nothing was found. It means nothing found was strong enough to cite, and the tutor answers from general knowledge without dressing it up as a quotation. Those thresholds are measured against the actual books, and tuned to fail in the safe direction: better to miss a citation that was available than to invent one that was not.
In Exam mode the line is drawn harder still. A question the curriculum does not cover never reaches the model — the request stops before any AI is called, and the student gets a plain note that it falls outside the syllabus. There is no answer to be careless with, because none is generated.
One path, so every feature answers the same way
The tutor is the visible case, not a special one. Lesson plans, worksheets, presentations and exam papers all reach the curriculum through the same single piece of code, and all come back with citations. When each feature searched its own way, none passed a grade or a subject — so a grade-3 worksheet could be built on a grade-10 chapter. Filtering by grade and subject before similarity is considered at all is the difference between finding "the book" and finding "a book".
The edition your school actually teaches from
Schools do not all teach the same material, and the national books are a floor rather than the whole of it. A school can add its own approved notes and chapters, and where its passage matches as well as a national one, the school's own wins. A school that prefers to work only from its own shelf can switch the shared corpus off. One school's uploads are never visible to another.
Why this matters most in Nepali and Social Studies
For a mathematics identity, a general model and the Nepali textbook mostly agree. For Nepali literature, for Social Studies, for the framing a Nepali board expects on a topic, a capable international model is often fluent, plausible and wrong — wrong about which text a grade studies, wrong about the emphasis, in a way that reads perfectly well to a student with no way to tell. Grounding the answer in the book that class was issued, and printing the page beside it, turns a risk a teacher must police into something a student can settle by looking.
None of this makes the system incapable of error. It makes its errors findable — the one property that lets a teacher put a tool like this in front of a class and stay responsible for what it says.