Source-grounded, not a chatbot: how we assemble study material

One source in, many views out

Give Novus Learn a single source — a Wikipedia revision or an uploaded document — and it does the extraction work: it splits the text into stable sections and source blocks, identifies direct claims, names the concepts and the labelled relationships between them, and measures coverage.

Every study view is then assembled from that structured, source-tied data. The summary is made of real sentences. The quiz questions are backed by real claims. Nothing is invented to fill a gap.

The word doing the work in that sentence is “assembled”. There is no generation step, so there is no point at which fluency can be substituted for accuracy. The pipeline's job is selection, ranking and arrangement — deciding what matters most and what shape to show it in — and those are decisions you can audit by following any item back to its claim.

What extraction actually produces

The intermediate structure is worth describing, because everything you see is a projection of it rather than a document in its own right:

• Source blocks — the text split into stable, addressable pieces, so a reference can point at one part of the source and keep pointing there.
• Claims — the direct statements those blocks make, each with the evidence text it came from and a score for how central it is.
• Entities — the concepts, people, places, organisations, events, dates, processes, stages and components the claims name, each citing the claims that named it.
• Relationships — typed, directional links between entities, marked by how directly the source supports them.
• Coverage — a measure of how much of the source made it into the structure, so a poor extraction is visible rather than silent.

How a claim becomes a card

Take a cloze card, the simplest case. The builder finds a claim, finds a term the claim mentions, and replaces the first whole-word occurrence of that term with a blank. The answer is whatever the source actually wrote, preserving its own capitalisation. There is no rewriting step in which the sentence could drift.

Quiz questions work the same way in reverse. The correct answer is a real entity the source names, and the wrong answers are other real entities of the same type from the same project — not plausible-sounding inventions. Option order is derived from a stable hash of the question and the option label, so the correct answer isn't always in the same position but the quiz still renders identically every time you open it.

Validators that refuse to invent

Under the hood, a validation pass rejects any claim, concept, or diagram node that the source can't back. Every reference in a finished study set has to resolve against the project it came from; anything that doesn't is removed before you ever see it.

If a topic doesn't provide enough claim-backed terms for safe flashcards, Novus Learn tells you that honestly instead of fabricating cards. The same rule governs the visual planner: a diagram is only recommended when the source structure genuinely supports that shape, and a node that would need an invented value doesn't get drawn.

That's a deliberate trade: fewer confident-sounding guesses, in exchange for study material you can actually trust and cite.

Where we are extra careful

Some material carries consequences if it is wrong. Step-by-step explainers are scanned for high-stakes signals — medical, legal, financial, emergency and hazardous content — and anything that matches is flagged and can never be presented as human-reviewed, whatever else happens to it.

Explainers are also verbatim by rule. A how-to is built from the source's own list items in the source's own words, never reworded into something snappier. If the source doesn't contain a real procedure, you get an honest “how-to unavailable” rather than an invented set of steps.

Ask a question, get a sourced answer

You can ask a topic a question in your own words. The question is classified into an intent and a set of focus terms, then answered by retrieving and ranking claims the project already holds. What you read back is real source sentences with evidence links — and if the source doesn't answer it, you get an honest “not answered” rather than a made-up reply.

Comparisons work this way too: ask how two things differ and you get each side's own claims placed next to each other, rather than a synthesised verdict about which is better.

The honest limitations

Extraction is not comprehension. Novus Learn works from sentence structure and repeated terminology, so a source that buries its point in a long subordinate clause, or names the same idea three different ways, will produce a weaker set than a plainly-written one. Tables, heavy markup and image captions are largely invisible to it.

It also cannot simplify. If the source is written for specialists, the study material is written for specialists, because the study material is the source. When that's the problem you have, the fix is a different source rather than a different setting — which is part of why uploading your own document is a first-class path rather than an afterthought.