Built for deep research
Long sources, questions, and deliverables — all in one workspace where you can verify the original.
Transformer compute: why do two studies disagree?
Sources
3 sourcesAttention versus recurrent layers
This appendix uses the per-layer complexity comparison from the paper “Attention Is All You Need” to make a rough estimate. It was written for this demonstration.
An attention layer’s computation grows with length² × dimension and a recurrent layer’s with length × dimension². Both studies use a model dimension of 512. Figure 2 plots this estimate for both layer types across input lengths and marks the two studies’ input lengths.
- Attention layer
- Recurrent layer
This is an order-of-magnitude estimate. It leaves out constant factors and the rest of the model, such as the feed-forward layers, so it cannot reproduce either study’s measured ratio.
Current source: Attention compute appendix (sample).pdf · Page 2 · Figure 2
Ask your sources
Result preview
After you ask, findings can be turned into a report.
Turn what you read into a perspective of your own.
Bring all your project’s documents together. Clarify your thinking through questions, comparisons, and verification. Shape your understanding into a report, so discussions and research are grounded in evidence and lead to concrete results.
Bring your sources
- Papers & industry reportsPDF
- Interviews & meetingsMP3
- Drafts & research notesWord · DOCX
- Data & spreadsheetsExcel · XLSX
Bring your questions. Together, we explore, compare, and check the evidence to build understanding.
Take your work further
- Word documentsDOCX
- PDF documentsPDF
- PPT presentationsPPTX
- Code filesHTML · Python
Your next piece of work,from the sources at hand.
A few reports. One recording. A decision ahead.
Connect the evidence. Give every judgment a source.
PDFArea foot traffic study.pdf01
“The visitor forecast assumes conditions after the rail line opens.”
Page 18 · Forecast assumptions
MP3Project discussion.mp302
“Phase one is due to reopen next year, but transport links may not be ready.”
Meeting excerpt · 12:48
Research report
Draft · To verifyShopping street renewal · Feasibility review
Frequently asked questions
What is DeepShelf?
DeepShelf is a cloud workspace designed for deep research and learning. Bring long sources such as papers, reports, recordings, drafts, and spreadsheets into one project, then ask AI questions, compare sources, and follow up. Turn your findings into downloadable reports. Citations in answers take you back to the original source with one click so you can verify them. Alongside DeepShelf 1.0, it supports GPT, Claude, DeepSeek, Kimi, and other models. Available models depend on your plan.
See models on each planHow is DeepShelf different from tools like ChatGPT? Why upload files first?
Tools like ChatGPT can read the files you send them, too. The difference is what happens after you upload. DeepShelf gives each source a first read and takes its own notes, a process called parsing. This has three benefits:
- Later answers draw on your sources first and are more accurate.
- Your sources stay together in one project, so you don’t have to send them again when you start a new chat.
- Every citation in an answer can be opened to check the original: a PDF opens at the passage in its original layout, a recording jumps to the matching time and transcript, and Word, TXT, and EPUB files open at the parsed text.
Can DeepShelf handle several long books or a stack of reports at once, and find how they relate?
Yes. Put them in one project, select the sources you want for this question, and ask where they agree, where they conflict, and how they connect. We’ve tested this with several full-length books at once, including questions that span books, and the results far exceeded similar products. That scale isn’t the ceiling, either. The real limits come from your plan: how many sources a project can hold and how large each file can be. Scans and spreadsheets don’t work as well as PDFs with selectable text. Open the citations to confirm each connection it points out.
What file types can I upload to DeepShelf?
DeepShelf supports PDF, Word (DOCX), plain text (TXT), ebooks (EPUB), images (JPG, JPEG, PNG), audio (WAV, MP3, M4A, OGG), webpages (HTML, HTM), spreadsheets (CSV, XLSX), PowerPoint (PPTX), and more. File size and the number of files per project vary by plan.
See supported source formatsCan I use DeepShelf for free?
Yes. New accounts get a seven-day Plus trial with 100 credits in total. When it ends, your account moves to Free automatically. Plus benefits end, but you can keep using DeepShelf. Parsing sources, asking questions, and creating reports all use credits, and how many depends on the length of your sources and the model you choose. Files you uploaded during your trial or a paid plan can still be asked about and cited on Free. Check your balance and history under Settings → Usage.
Compare plan limitsWill the papers, recordings, or client materials I upload to DeepShelf be used to train models?
DeepShelf does not use your uploaded sources, conversations, or generated results to train large language models.
We keep your sources, along with the generation process and results, long-term so you can revisit them and continue working at any time. When you no longer need them, you can delete projects or sources in the workspace. We also keep the request logs needed to diagnose faults, maintain security, and prevent abuse.
DeepShelf is a cloud service. To parse files and generate answers, the content each task needs is sent to the relevant model and parsing services. We recommend not uploading sensitive information unrelated to the current task.