Upload your sources.
Ask anything.
Drop in a PDF, a link, a YouTube lecture, or your notes. Every answer comes from exactly what you attached — click a citation to read the passage it came from. And if it isn't in your sources, it says so.
free to start · Google sign-in · answers only from your sources
Agentic RAG — A Survey
arxiv.org · 42k chars
Attention — full lecture
youtube.com · 48 min
lease_agreement_2026.pdf
local · 14 pages
Notes — hybrid retrieval
pasted · 3.1k chars
3 ready · 1 processing
artifacts save to your account
features
Grounded by design.
Everything below follows from one rule: answers come from your sources, or they don't come at all.
Citations that open
Every answer cites its sources inline. Click a [n] to jump to the exact passage — or hover to read it right there.
It tells you when it doesn't know
Ask something outside your sources and you get a plain “not in your sources” — not a confident guess dressed up as an answer.
Studio, one click
Turn the same sources into a mindmap, quiz, flashcards, summary, fact-check, or a cited research report. No re-uploading.
Sign in with Google
One click, no passwords. Your notebooks, sources, chats, and artifacts follow you across devices.
Three panes, your layout
Sources, chat, and studio side by side. Drag to resize, collapse what you don't need — the layout sticks.
Private by default
Every notebook is scoped to its owner at the database level — not just hidden in the UI. Your corpus stays yours.
how it works
Drop it in. Ask away.
Four steps between a file on your desktop and an answer you can verify.
extract
Attach
Add a URL, PDF, YouTube video, or pasted text to a notebook. Whatever you want to work with.
chunk → embed
Index
Text is extracted, chunked, and embedded in the background — you can start asking before it finishes.
retrieve → rerank
Ask
Hybrid retrieval and reranking pick the passages that matter, then the model answers from those. Only those.
gate → cite
Check
Answers cite passages you can open. Nothing relevant found? No model call, no answer — the guess never runs.
studio
One click. Four artifacts.
Same sources, same citations — turned into a mindmap, a quiz, a fact-check table, or a full research report.
Which mechanism lets the model attend to multiple representation subspaces at once?
explanation: Q / K / V are projected into h parallel subspaces [2]
the no-bluff check
Try to trip it up.
Ask about anything you didn't upload. The answer is a plain refusal — enforced in the pipeline, not added as a disclaimer.
· retrieval → 0 passages cleared the gate
· model call → skipped · 0 tokens
→ “I don't know about this.”
· retrieval → top-k 6 · rerank 6 · gate passed
· streamed answer · citations [2][4]
grounded ✓
the system prompt, literally
“You answer only from the provided passages. Cite [n] for every claim. If the passages do not contain the answer — even partially — reply exactly ‘Nothing related to this is stated in the sources’. Never answer from your own knowledge.”
Relevance gate
If no retrieved passage clears the bar, the model is never called. There is nothing to hallucinate from.
Citation guardrail
Every [n] is checked against what was actually retrieved. Citations that don't resolve get stripped or regenerated.
Tested, not claimed
“Does it refuse every out-of-context question?” is part of the eval suite — and it gates merges.
get started
Open a notebook.
Sign in with Google, paste a link, ask your first question. Every answer you get will be one your sources can back up.
one-click Google sign-in · free to start