AI quiz generator built on your own material
Not generic questions — questions from the exact chapter you just read, marked by something that tells you what you missed.
Quiz me freeFrom a chapter you have read to questions you cannot bluff
Rereading a chapter feels like studying and mostly is not. You recognise every line, so you conclude you know it, and then the exam asks you to produce it from nothing. The fix is to be asked — but the questions have to be about your material, and somebody has to tell you what your answer actually missed. That is the whole of this page.
1. Upload the material you are being tested on
The textbook chapter, the lecture deck, the standards document, the problem set, a photo of the page. PDF, DOCX, PPTX, Markdown, plain text, or an image — up to 30MB a file. Tali reads it and breaks it into short notes, one idea each, and the questions are generated per note. That is why a question can always be traced back to a passage you uploaded.
2. Get questions off your own pages, note by note
For each note Tali writes four to six fill-in-the-blanks — a sentence from your own chapter with the load-bearing term taken out — plus multiple-choice questions with four or five options where more than one can be right, so elimination does not get you there. Across the product there are four question types: single choice, multiple choice, fill in the blank, and true or false. Then there are the open-ended ones, where you have to write the reasoning out.
3. Answer, get marked, and meet what you missed again
Choice and blank-filling questions are marked instantly against the key, with the explanation that was written alongside the question. Open-ended answers go to an AI grader, which reads your working against the reference answer and comes back with a verdict, a score out of 100, and a paragraph naming the step you got wrong. Get one wrong and you can ask for a variant — a new question on the same idea — and answer that one instead.
Why this beats a question bank you found online
Every question comes off a page you uploaded
A bank of practice questions for organic chemistry is written against somebody else's course. It half-overlaps yours: a third of it is material your lecturer skipped, and the topic your exam leans on is the topic it happens to be thin on. Tali generates each question from the text of one of your notes, which came from one of your files, and it keeps the link — every question can show you the note it was written from.
It marks the open-ended ones too
A multiple-choice question can be guessed. An answer you have to write out cannot. Tali passes your working, the question, and the reference answer to a grader that returns three things: whether you were right, a score out of 100, and prose feedback that says which step broke and what your answer never mentioned. Having an answer marked this way costs tokens: being told exactly where your proof broke is a real model call, not a lookup in an answer key.
Getting one wrong is the useful part
Every question has a variant button. Tali writes a fresh question on the same concept — same idea, different framing — and asks you that one, so you cannot pass by remembering that the answer was C. Generating a variant is generating a question, so it costs tokens; that is worth knowing before you lean on it. Answer the variant correctly and the original question moves forward in your review schedule. You graduated it by answering, not by telling yourself you knew it.
The session asks the right questions, in the right order
Start a practice session and Tali does not shuffle a pile. It puts what is due for review first, then the questions you have previously got wrong, then ones you have never seen — and it leaves out anything you answered correctly in the last three days, because being asked the same thing again on Thursday is not practice. The session is sized to whatever is left of your daily goal, so it ends when you are done rather than when the file runs out.
Who ends up quizzing themselves here
The chapter you have read three times
You know every paragraph by sight and could not reproduce one of them under pressure. Upload it, answer the questions it generates, and in ten minutes you know exactly which parts you have and which parts you were only nodding along to.
Problem sets whose answer key is a single number
The book says the answer is 3.2% and stops there. It never tells you that your second step was wrong and the third one cancelled the error out. Write your working into an open-ended question and Tali marks the working, not just the final line — which is the only feedback that is any use on a proof or a derivation.
The topic you keep getting wrong
Tali tracks how you score per knowledge point across everything you have answered. When some of them keep coming up short, you can have it generate a fresh round of practice aimed only at the weak ones — rather than another undifferentiated pass over a chapter you mostly know.
Questions people ask before they upload anything
You get 600 free tokens when you create an account, no card required. Generating a practice question costs 10 tokens, so that balance is worth roughly sixty questions before you spend anything. Answering a multiple-choice, fill-in-the-blank or true-or-false question is free, however many times you come back to it. Having an AI mark an open-ended answer costs 10 tokens, and so does asking for a fresh variant of a question — both are real model calls, and we would rather say so than bury it. There is no subscription; you top up when you want more.
Four types: single choice, multiple choice (four or five options, sometimes with more than one correct answer), fill in the blank, and true or false. On top of those are open-ended questions — the ones where you type out the reasoning and Tali marks the reasoning. From a single note it aims for four to six fill-in-the-blanks and a couple of multiple-choice questions, so a chapter of a dozen notes comes back as dozens of questions, not five.
Your answer goes to the model together with the question and the reference answer, and comes back as a verdict, a score out of 100, and written feedback that names the step where you went wrong. It is the least comfortable feature in Tali, and the one people keep. If the grading call fails — bad connection, mangled response — you can ask for it again.
They are written from the file you uploaded rather than from whatever a model half-remembers about the subject, so they stay on your syllabus and use your material's terminology. It is still AI: it can misread a table, or ask something the chapter never quite claimed. Every question carries its explanation, and there is a button that makes Tali walk through the reasoning. If a question looks wrong, check it against the note it came from — that note is one click away, and it sits next to your original file.
PDF, DOCX, PPTX, Markdown, plain text, and photos or screenshots of pages (PNG, JPG, WEBP, HEIC). Up to 30MB per file. A whole textbook is better uploaded a chapter at a time — the questions come back tighter, and you were going to study it a chapter at a time anyway.
It joins your review queue automatically, right or wrong, and comes back on a spaced schedule — sooner if you missed it, much later if it was easy. There is nothing to enrol and no deck to maintain; answering the question once is the enrolment. Reviewing what is already in the queue does not cost tokens.