Methodology

How Cognia AI works

Most AI study tools sit on top of a general chatbot. Cognia AI is the opposite: every answer the tutor gives is grounded in your own uploaded material, with a citation to the exact slide, page or lecture timestamp it came from. This page explains, in plain language, how that works.

The problem with general chatbots for studying

ChatGPT, Claude and Gemini are trained on the public web. They are excellent at explaining concepts in general. They are bad at knowing what your lecturer specifically said in week 7. When you ask them about your course, they generate a plausible-sounding answer that may or may not match what your unit teaches. That is fine for casual reading. It is dangerous for exam prep.

Cognia AI was built to remove that danger.

Step 1 — You bring the source

Studying with Cognia AI starts with you uploading your own material:

• A live or recorded lecture (audio)
• A PDF (textbook chapter, paper, slide deck, scanned handout)
• Your own notes (typed or pasted)

Audio is transcribed with timestamps. Documents are parsed for structure (headings, sections, lists, figures). Both end up as text plus structural metadata.

Step 2 — The source becomes a searchable knowledge base

Cognia AI splits your material into small overlapping passages (commonly called “chunks”) and converts each one into a vector embedding. Every chunk knows which source it came from and where in that source it lives — slide number, page number, lecture timestamp. This is the index that powers everything downstream.

Step 3 — When you ask, the tutor retrieves before it answers

When you ask the tutor a question, the system does three things in sequence:

1. Embeds your question and finds the most relevant chunks of your material.
2. Passes those chunks plus your question to the language model as grounded context, with explicit instructions to answer only from that context and to cite which chunk each claim came from.
3. Renders the model's response with the citations as clickable references back to the source.

This pattern is sometimes called retrieval-augmented generation, or RAG. The version that matters for studying is the strict version: the model is not allowed to fall back to its training data when your material does not cover the question. If the tutor cannot answer from your sources, it tells you so rather than guessing.

Step 4 — The other study modes use the same index

Flashcards, quizzes and podcasts are all generated from the same chunked knowledge base.

• Flashcards are written by extracting the high-recall facts and definitions in each chunk and pairing them as Q/A.
• Quizzes generate plausible distractors from adjacent chunks (so wrong answers feel like real misconceptions you might hold), then mark the result against the source of truth in your material.
• Podcasts turn the structured outline of a source into a two-host conversation, with each segment tied back to a section of the source it came from.

What this means in practice

Three properties fall out of this design:

1. Verifiability. Every tutor answer has a citation you can click in two seconds. If the citation doesn't support the claim, you know the model is wrong and can flag it.
2. No drift. The tutor cannot wander into confidently-wrong general knowledge because the prompt only includes your sources.
3. Course alignment. Flashcards and quizzes are generated from the exact lecturer voice and notation your unit uses, not from a generic textbook on the same topic.

What this means for academic integrity

Cognia AI is a study aid, not a cheat sheet. It helps you understand and remember your own material faster. It does not write essays for you and we strongly discourage submitting AI-generated work as your own. The citation-first design is there precisely so you can use the tutor as a study partner you can verify against, rather than a source you copy from.

For the practical version of this, see our guide to using AI for uni exam prep without cheating.

What we don't do

A few things explicitly out of scope by design:

• We do not train shared models on your content. Your uploads stay yours.
• We do not sell your data. Stripe handles billing so we never see card details.
• We do not generate practice or mock exams from external syllabi we do not have permission for. Quizzes come from your own material only.

Try it on your own course

The fastest way to see if the methodology holds up is to upload one lecture or PDF you know well, ask the tutor a question you already know the answer to, and check whether the citation actually supports the claim. If it does, you can trust it on the questions you don't already know the answer to.