RAG over textbooks — why citations matter more than the answer

A confident wrong answer is the worst kind

General-purpose chatbots are built to always produce an answer. In a search box, that's fine. In a classroom, it's a liability. Ask a generic model about the water cycle and it returns a fluent, plausible, textbook-shaped paragraph — that may not match the board your school follows, the edition of the book on the student's desk, or the chapter the teacher actually taught.

Worse, it will invent a reference: "see page 142." There is no page 142. The student copies it into a notebook. The teacher spots it. And trust doesn't come back after the first fabricated citation.

The instinct is to fix this with a "smarter" model. It isn't a model problem. A model can only answer from what it absorbed in training — a blurred average of the public internet. It has never seen your school's textbook.

What "trained on your textbooks" actually means

This is where retrieval-augmented generation — RAG — comes in, and it's worth being precise, because the phrase gets used loosely.

Teach Smart doesn't fine-tune a model on your books. It indexes them. Every textbook, workbook, and piece of additional material a school uploads is split into small passages and converted into vectors — 1,536-dimension embeddings — and stored. When a question comes in, the system first retrieves the handful of passages that genuinely match it, then asks the model to answer only from those passages.

The model still writes the sentence. But the facts come from the book on the desk — not the book in its memory.

Citations are the product, not a footnote

If an answer can't tell you which chapter, page, and paragraph it came from, it isn't grounded — it's a guess with good grammar.

Every Teach Smart answer carries a citation down to chapter, page, and paragraph. We treat that as the main feature, not a polite footnote, for three reasons:

• A teacher can verify it in five seconds. Open the page, read the paragraph, done. Verification that takes five seconds gets done; verification that takes five minutes doesn't.
• It teaches students where to look. A cited answer doesn't just hand over a fact — it points at the source, so the student learns the book, not only the answer.
• It's aligned by construction. Because the answer is built from the school's own indexed material, it automatically matches the right board, the right edition, the right syllabus. No "which curriculum?" guessing.

Why a good RAG system refuses to answer

Here's the part that surprises people: a well-built RAG system should sometimes say nothing.

If retrieval comes back empty — the question isn't covered anywhere in the indexed material — the honest output is "this isn't in the material I've been given," not a confident paragraph assembled from the model's general training. A guess dressed as a citation is exactly the failure we set out to remove.

Refusing to answer is a feature. It tells the teacher: this topic isn't in the uploaded books — either it's out of syllabus, or the reference material needs to be added. Both are useful to know. A chatbot that never says "I don't know" can never be trusted when it does answer.

What it changes — for teachers and students

For teachers: Textbook Chat grounded in the school's own books is something you can hand to a class unsupervised. The worst case isn't a wrong answer spreading through the room — it's a cited answer you can check, or an honest "not in the material." That's a tool you can trust enough to step away from.

For students: the citation turns a chatbot from an answer vending machine into a study guide. "Where did this come from?" has an answer every time — and following that citation back to the page is the actual studying.