We will never let our AI hallucinate your textbook

What a hallucination actually is

A large language model is a probability engine. Given some text, it predicts the most plausible next token, then the next, and so on. It is extraordinarily good at producing text that sounds right. Crucially, it has no built-in concept of whether the text isright — plausibility and truth are different things that usually, but not always, coincide.

A “hallucination” is what we call it when they diverge: the model generates something fluent, confident, and false. Ask a raw model for the date of a treaty or the value of a constant and it will often answer correctly — and occasionally invent a number with exactly the same confidence. For casual use that’s annoying. For a student memorizing that number for an exam, it’s a trap.

Grounding: answer from sources, not from memory

The single most effective defense against hallucination is to stop asking the model to recall facts at all. Instead of “What does the model remember about mitochondria?”, we ask “Given these specific passages from the student’s notes, answer the question and cite them.” This pattern is called retrieval-augmented generation (RAG), and it changes the model’s job from recalling to reading.

The pipeline behind a single tutor answer looks like this:

1. Index.When you add a note, PDF, or lecture, we split it into small overlapping chunks and compute an embedding — a numerical fingerprint of meaning — for each one. These live in a vector index scoped to your account.
2. Retrieve. Your question gets the same embedding treatment. We find the chunks whose meaning is closest to the question, then re-rank them so the most relevant passages rise to the top.
3. Generate.Only those passages — not the whole library, and not the open internet — go into the prompt, with an instruction to answer strictly from them and attach a citation to each claim.
4. Attribute. We map each cited span back to the exact source chunk so the answer renders with a clickable reference to the line in your notes it came from.

Because the model is reading supplied text rather than reaching into its weights, the failure mode shifts from “invents a plausible fact” to “summarizes the wrong passage” — and the second one you can catch, because the citation is right there to check.

Why retrieval quality is the whole game

Grounding only helps if the right passages get retrieved. Feed the model irrelevant chunks and it will dutifully ground its answer in the wrong thing. So most of our engineering effort goes not into the generation step but into retrieval:

• Chunking that respects structure.We split on semantic boundaries — headings, paragraphs, slide breaks — not arbitrary character counts, so a chunk is a coherent idea rather than half of two.
• Hybrid search.Pure semantic search misses exact terms (a specific theorem name, a date); keyword search misses paraphrases. We blend both so “the powerhouse of the cell” and “mitochondrial ATP synthesis” can find each other.
• Re-ranking. A first pass casts a wide net; a second, stricter model re-orders the candidates by true relevance to the question before anything reaches the prompt.
• Scoping. Retrieval is confined to yourlibrary by default, so answers reflect what you’re actually studying — your professor’s framing, not a generic one from the internet.

The most important feature: refusing to answer

Here is the part that separates a study tool from a chatbot. When retrieval comes back empty — you asked about something that isn’t in your notes — the right answer is not to improvise. It’s to say so.

This is genuinely hard to ship, because a refusal feels worse in the moment than a confident answer. A model that always answers demos better. But for the one student who would have memorized the fabricated answer, the refusal is the feature. We’d rather say “I don’t know” a hundred times than be confidently wrong once.

The honest limits

No system is perfect, and claiming otherwise would be its own kind of hallucination. Grounded generation dramatically reduces fabricated facts, but it can still:

• Misread a correct source— summarize a retrieved passage slightly wrong. The citation lets you catch it; we surface the source prominently for exactly this reason.
• Inherit errors in your notes. If your source says something wrong, a faithful answer repeats it. We ground in yourmaterial; we don’t fact- check your professor.
• Retrieve an adjacent-but-wrong passage for ambiguous questions, which is why re-ranking and the relevance threshold matter so much.

Our job is to push these toward zero and to make the remaining cases visiblerather than invisible. That’s the design principle under everything here: a wrong answer you can check beats a wrong answer you can’t. Every citation is an invitation to verify, and we think that’s what trust in an AI tutor actually looks like.

Want to see grounding in action? Add a few of your notesand ask the tutor something only your material would know — then click the citation.

Frequently asked questions