How it works
How Paper turns your sources
into a study partner.
Upload a PDF, a lecture recording, or a website. Paper indexes the content, then lets you chat with it and study it - with every answer tied back to a specific page or timestamp.
The pipeline
From upload to cited answer in four steps.
Paper is built on retrieval-augmented generation (RAG) - the pattern that lets an AI model answer questions about your own documents instead of guessing from its training data. It looks something like this:
- 01
Upload a source
PDF, audio, video, EPUB, Word, PowerPoint, web URL, YouTube link, or an import from Notion or Outline. Paper handles the extraction.
- 02
Paper extracts the text
Scanned PDFs run through OCR. Videos and audio are transcribed (with speaker labels). The full text is split into passages on natural boundaries - paragraphs, page breaks, timestamp groups.
- 03
Each passage gets an embedding
An embedding is a numerical fingerprint that captures the meaning of the passage. Passages with similar meaning end up close together in the index, even if they use different words.
- 04
You ask a question. Paper finds the right passages.
When you chat, Paper searches the index for the passages most relevant to your question, then asks an AI model to answer from that context, with citations back to the source. When source-only mode is enabled, the model is instructed to answer only from your selected sources.
Trust
Source-grounded answers cite their source.
Paper is designed to cite the page or timestamp behind its answers wherever the answer depends on your sources. You see something like [p. 5] or [3:42] beside each claim. Clicking it jumps you to the exact spot in the PDF viewer or the video player.
This is the simplest defence against hallucination: every important claim should be traceable to something real. If an answer seems off, you can open the citation and check the original source in one click.
Control
You decide what the AI sees.
Each chat sits on a page with one or more sources attached. You can:
- Toggle individual sources on and off for any chat turn.
- Narrow a PDF to a specific page range (e.g. only chapters 4 to 7).
- Narrow a video or audio source to a specific timestamp window.
The AI only ever sees what's selected. Pages don't bleed into each other, and other users' content is never reachable from your chat - see Security for how we enforce that.
Models
Three AI models. One click to switch.
You can switch models from the chat settings menu. All three are cloud-hosted; no AI runs on your machine.
| Model | Provider | Best for |
|---|---|---|
| Grok (default) | xAI | Fast chat and everyday study |
| GPT | OpenAI | Harder reasoning and structured answers |
| Gemini | Long context and web-aware tasks |
Runtime
All cloud. Nothing to install.
Paper runs in your browser. There's no desktop app, no local model, no Python environment. When you ask a question, your prompt and the retrieved passages are sent over HTTPS to Paper's servers, which call the AI provider you've selected. Your sources stay in Paper's database the whole time.
Cost
Free during beta, with a free tier after launch.
You don't bring an API key - Paper pays for every model call. Paper is free during beta, and we plan to keep a generous free tier after launch. Paid tiers may be introduced later for heavy use, very large libraries, and team workspaces.
Under the hood
A few questions we get asked a lot.
How does retrieval handle different kinds of questions?
How do you stop the AI from confidently quoting the wrong passage?
What happens on a really big document?
Why citations on every claim?
Try it
Bring your own PDF.
Upload a textbook, a lecture recording, or a research paper. Free during beta, no credit card.