Revising from lecture slides, textbook chapters and past papers
Course material can be queried without uploading anything: up to 40 files, 25 MB each, parsed and indexed inside the browser tab. This page shows what to load for one term, how to ask questions that lead back to slide and page numbers, and which uses cross the line your institution draws.
Last updated:
A term of material is usually 15–30 files, which fits the 40-file cap once. So the real decision is which files to include, not how to squeeze them in.
What to load for one term
| What you have | Format | Worth importing? | Note |
|---|---|---|---|
| Lecture slides | PDF export | Yes — text and slide numbers come through | A slide that is one big image keeps only its caption |
| Textbook chapters | PDF with a text layer | Yes | Scan-only editions yield nothing at all |
| Past papers | DOCX or PDF | Yes, all of them | Often the highest-value set for spotting repeated question patterns |
| Your own notes | Markdown or TXT | Yes | Your wording helps retrieval, because you ask questions in your own vocabulary |
| A scanned textbook | Image-only PDF | No, not before OCR | The parser reports the file instead of indexing emptiness |
| Marks in a spreadsheet | CSV after export | Rarely useful | Retrieval is text-based; numbers alone have little context to match |
A revision loop built on citations
- Load the term’s material into one library and let indexing finish; progress is shown per file.
- Ask a definition question you can already answer, such as “How does this course define opportunity cost?” Check that the citation lands on the right slide.
- Ask a gap question: “Which week 3 topics appear in the 2024 paper?” Answer the retrieved questions from memory before checking the slides.
- Ask a comparison question across sources: “Do the slides and the textbook define variance differently?” Both passages come back, and you decide which one your examiner follows.
- Keep the citations you used in a Markdown file. Re-reading your own question list is faster than re-reading a chapter.
Where the material fights back
- Diagrams and timelines carry meaning a text pipeline cannot see, so a question about a diagram returns only caption text.
- Formula-heavy pages usually lose superscripts and subscripts, which changes what an expression says. Read the formula in the source.
- Handwritten notes in a scan do not parse at all.
- A 40-file library holds one term comfortably. A whole degree means clearing site data between terms, which also deletes the index — export what you want to keep first.
The 20,000-chunk ceiling, in context
Chunks run about 700 characters with 15% overlap, so 20,000 chunks is around 14 million characters of text. A 60-slide lecture PDF is often under 30,000 characters, which means the 40-file cap normally binds long before the chunk cap does.
When this is not the right approach
- You want a chapter-by-chapter summary of a whole book in one answer. Retrieval returns the top six passages, not a condensation.
- The work is graded and you were told to do it unaided. Read your course rules first: running locally does not change whether assistance is permitted.
- You need the same library on a phone and a laptop. The index sits in one browser profile and is not synchronised.
- You want to hand a study set to a group. Files never leave the device, so the index cannot be shared — share your question list and citations, or each person imports their own copies.
- Your device has little free memory. Mobile browsers cap collections lower, and the tool states the reduced limit instead of failing part-way through indexing.
Settings that suit studying
- Stay in the default retrieval tier: it quotes the retrieved sentences and keeps numbered citations, which is what you want to memorise from.
- Set strictness to documents only. When nothing clears the similarity threshold you get “not found in your documents” instead of a plausible invention.
- Turn on the generation tier only if you have WebGPU; without it, CPU runs at roughly 3–8 tokens per second and the download is 400 MB to 1.0 GB.
- Files never leave the device, there is no account and no tracking pixels, and the library works offline once the embeddings are cached.
Does it work on a phone?
Yes, with a lower collection cap because mobile browsers allow less memory. The interface reports the reduced limit rather than failing during indexing.
Will it summarise an entire textbook?
No. It retrieves the six most relevant passages for a question. For a chapter summary you would be reading a hallucination risk rather than a source.
Can my study group share one library?
Not directly. There is no server and no shared index; each person imports their own copies of the same files, and the index stays in their browser profile.
Is this cheating?
That is defined by your course, not by this page. A retrieval tool that quotes your own course material is closest to a search index of your notes; submitting generated prose as your own work is a different matter.
Why did my textbook PDF index as empty?
It is almost certainly an image-only scan without a text layer. The parser detects this and reports the file, leaving the rest of the library usable.