Loading…
Loading…
CiteSeek streams answers grounded in the files you upload — and every sentence links back to the passage it came from, so you never have to take the model's word for it.
No account. No card. Nothing to install.

Step 1
It is embedded and compared with the passages of the files in this workspace — scoped in the SQL of the search itself, not filtered afterward.
Step 2
Whatever clears the relevance floor goes to your browser first. The model then writes with numbered markers against that payload, so a marker points at a passage that already arrived.
Step 3
Clicking one opens the document at the exact characters the answer was grounded in, highlighted, with its page number — the offsets are recorded when the file is chunked.
Measured rather than asserted: the method behind each number, and what it does not cover, is in the README's Numbers section.
The passages are retrieved and sent to your browser before the model writes a word. A marker in the answer resolves against that payload, so it points at a passage that already existed — the model chooses which to cite, and cannot cite what it was not given.
A marker that resolves to nothing stays plain text rather than becoming a link, so a number the model invented cannot present itself as a source. Neither can a marker in a reply that opens in the form the model is told to use for declining: none of them becomes a link, and the reply says why, even when it then answers correctly, as can happen mid-conversation.
And when nothing retrieved is relevant enough, no answer is generated at all: the reply saying so is written by CiteSeek, not by the model.
Local mode runs the model on your machine, so the text of a document never leaves it — the one claim here that is sovereignty rather than a region on a map. It is experimental, and the page says what it measures.
A fictional company handbook, parsed into passages that kept their page numbers. Ask it something and follow a marker back to the line it came from.
Open the demo workspace