Answers you can check, line by line.
RAG Notebook reads your PDFs, videos, and web pages, then answers from them — and only from them. Every sentence links back to the page or timestamp it came from, so verifying an answer takes a click instead of an afternoon.
1,000,000 tokens, 10 workspaces, no card required.
Most RAG tools answer confidently. This one answers checkably.
A fluent paragraph with nothing underneath it is the failure mode of every retrieval tool — and it’s invisible until someone acts on it. The fix isn’t a better-worded prompt. It’s better retrieval, and evidence attached to the output.
So we made the evidence non-optional. If a claim can’t point at a passage in your sources, it doesn’t get made.
Accuracy is a pipeline, not a prompt
Three things happen between your question and the answer. Each one exists to throw away the wrong passage before it reaches the model.
Two searches, not one
Each question runs a dense semantic search and a sparse keyword search side by side — 40 candidates each. Meaning-based search alone misses exact terms; keyword search alone misses paraphrase. Running both catches what either would drop.
Fused, then narrowed hard
Reciprocal rank fusion merges both lists, and only the top 8 passages survive — at most 4 from any single source. A long document can’t crowd out the one paragraph that actually answers you.
Bound to its evidence
Every sentence carries the passage it came from. Click a marker and the PDF opens on that page with the exact text highlighted, or the video seeks to that second. Checking an answer takes one click, not a re-read.
Built for the moment someone asks “where does that come from?”
It says when it doesn’t know
When the retrieved passages don’t answer the question, you get told that — not a confident paragraph assembled from nothing. Silence is a valid answer here.
PDFs with real page anchors
Citations resolve to a page and a highlighted run of text in the rendered document, not a vague filename reference.
Video that answers by timestamp
Drop a video or a whole playlist. Transcripts are chunked on topic boundaries, so a citation lands on the moment, not the hour.
A workspace, told back as a podcast
Turn every ready source into a ~5-minute audio overview — two hosts talking through the material the way a briefing would. Generated on demand and streamed back as audio you can play anywhere.
Web pages, kept as sources
Add a URL and it becomes a first-class source in the workspace — retrieved, ranked, and cited like everything else.
Memories that carry over
Tell a workspace what matters once — house style, an acronym, who the audience is — and every later answer takes it into account.
Indexing status you can trust
A source reads “indexed” only after the server finishes embedding it. No progress bar theatre, no answering from a half-read file.
A playlist arrives as a course, not a pile of videos
Add a YouTube playlist and the server lays out a route through it once indexing finishes: modules in order, each one saying what you should be able to do at the end of it, how long it runs, and which videos it covers.
- Up to 20 videos in a playlist on the free plan, 200 on Pro.
- Every video in the route stays a source in the workspace, so a question about the course comes back cited to the minute it was answered at.
- Videos that never finished indexing are named, not hidden — you always know what the route doesn’t cover.
- Ticking a module off is kept in your browser, so the route remembers where you stopped.
A free plan you can actually work in
Most tools hand you three documents and a countdown. You get 10 workspaces, 7 sources in each, and 1,000,000 tokens — enough to finish a real project before you decide anything.
| Limit | Free | Pro |
|---|---|---|
| Workspaces | 10 | 100 |
| Sources per workspace | 7 | Unlimited |
| Tokens included | 1,000,000 | Unlimited |
| Upload size | 10 MB | 100 MB |
| Videos per playlist | 20 | 200 |
| Podcast overviews | 1 | 10 |
| Citations on every answer | Included | Included |
Your documents, and nobody else’s business
Your sources stay yours
Documents are stored per workspace, scoped to your account, and never pooled into training data.
Authenticated on every call
Every request carries a signed session. There is no anonymous read path to a workspace, ever.
Retrieval runs server-side
Embeddings and ranking happen on our infrastructure and answers are streamed back — your files are never shipped to the browser to be searched.
Delete means deleted
Remove a source and its chunks and embeddings go with it. Delete a workspace and the whole index is dropped.
Questions worth asking
What stops it from making things up?
Answers are generated only from passages retrieved out of your own sources, and every sentence is bound to the passage behind it. If retrieval comes back empty, the answer says so instead of filling the gap. You never have to take a claim on faith — the evidence is one click away.
Is the free plan a trial?
No. It’s a plan. You get 10 workspaces, 7 sources in each, and 1,000,000 tokens without entering a card. Upgrade when you outgrow it, not when a countdown ends.
What can I put in a workspace?
PDFs and documents up to 10 MB on the free plan, videos and playlists up to 20 videos, and web pages by URL. Everything in a workspace is searched together, and every answer tells you which source it leaned on.
Do you train on my documents?
No. Your sources are used to answer your questions inside your workspaces. They are not pooled, not shared between accounts, and not used as training data.
What happens when I run out of tokens?
Answering pauses and the app tells you plainly, with your remaining balance shown in your profile. Nothing is deleted, and your workspaces stay exactly as you left them.
Put your sources in. Get answers you can defend.
Create a workspace, add a PDF, and ask it something you already know the answer to. Then check the citation.