Citation provenance
Every claim points to a page you can open
Agent Bayes retrieves from the papers you indexed before it writes a single line. Every claim in your mindmap carries the author, the year, the printed page range, and the passage it was drawn from. Click one and the PDF opens at that page.
Free to start, no card. 200 credits when you verify your email.
The reason you are here
You have already checked a reference that did not exist
The author was real. The journal was real. The year was plausible. The paper was not. Maybe you caught it before a supervisor did, and maybe you spent the next hour checking the other eleven references in the same output.
After that, the tool stops saving you anything. Verifying every line by hand costs more than writing the paragraph yourself, so the output goes in a drawer and you go back to the stack of PDFs.
The fix is not a better-behaved model. The fix is a system that retrieves from your library first and keeps the receipt attached to what it wrote.
How the grounding works
Retrieval first, then writing, in that order
Three stages, and the order is the whole point. Nothing reaches your mindmap that did not come back from your own corpus.
You decide what it can read
Upload PDFs, or include items straight from Zotero. Papers are indexed into knowledge bases you attach to a project, and that set is the only thing retrieval ever sees.
Retrieval runs before writing
Around ten queries per task run in parallel across your corpus and are re-ranked together. The stage that writes into your mindmap only sees passages that actually came back.
Each claim keeps its evidence
Author, year, printed page range, and the exact chunk of source text stay attached to the node. Click a citation and the reader opens at that page.
In the product
The evidence sits one click from the claim
Citations carry their bibliographic detail and their page range. Node properties keep the provenance of every edit, by you or by the agent.
What you can rely on
Six things that make a claim checkable
Printed page numbers
Page references use the actual printed pages of the journal or book, not the PDF's internal count, so they go straight into your bibliography.
As many citations as the claim needs
A synthesis drawing on four papers carries four citations. One representative source standing in for a body of work is how detail gets lost.
A confidence score on every node
Each claim is scored from 0 to 100 on source strength and agreement between sources, so a thin claim looks thin instead of reading like a settled fact.
Retrieval stays inside your corpus
The agent does not reach for outside information. A paper you did not add is a paper it cannot cite.
Gaps get reported, not filled
When your library does not support an answer, the system tells you that. Reporting the limit is a feature, not an error state.
The whole history stays attached
Which model wrote a node, which sources it drew on, what changed since, and whether you or the agent changed it.
What Agent Bayes does not do
- It is not a search engine. It reads the library you give it, so a paper you never added is a paper it cannot cite.
- It does not write your paper. The mindmap is material you argue from, and the argument stays yours.
- It does not make a language model infallible. What it does is make every claim traceable in one click, which is a different and more honest promise.
The last one is deliberate. A tool you cannot check is a tool you cannot defend, and defensibility is the entire point.
Before you ask
Questions researchers ask first
Which papers can it cite?
Only the ones in the knowledge bases attached to your project. You decide what goes in, and retrieval never reaches past that boundary.
How fast can I check a citation myself?
One click. The built-in reader opens the document at the cited page with the passage in view, and the extracted bullet points sit in a sidebar that follows your scrolling.
What if my sources are not in English?
Indexing writes structured summaries in English whatever the source language, so a French, German, or Japanese paper stays searchable and citable. You can also request an inline translation of any passage while reading.
Does it work with the library I already keep in Zotero?
Yes. A native plugin for Zotero 7, 8, and 9 adds items into a knowledge base from inside Zotero, with no export and no re-uploading, and it shows the estimated indexing cost before you commit.
What does it cost?
The free plan costs nothing and gives you 200 one-time credits once you verify your email, which covers about five indexed papers. After that, monthly plans are metered in credits: indexing an average 20 to 30 page PDF costs about 20 credits, and a typical research query costs 10 to 20. Estimated cost is reserved up front and only actual usage is charged.
Full detail lives in the documentation and the pricing page.
Related
If you are also wrestling with
Create your account
Put your own library behind every claim
The free plan indexes about five papers and gives you 200 credits once you verify your email. Enough to watch it cite your own sources and check whether the pages match.
- No credit card needed
- Retrieval never leaves your library
- Works with Zotero 7, 8, and 9
Want to look around first? Read the product tour or compare us against the tool you use now.