Le graphe

Under everything is one graph. It's worth understanding its parts, because every other feature is just a different view onto it.

Les nœuds sont des affirmations et des articles

A paper is a source you imported — a PDF or its metadata. A claim is a note you wrote in your own words off a span of a paper. Papers are what you own; claims are what you understand. Most of the graph's value lives in the claims, because those are the parts only you could have written.

Plein ou en creux

On a board, a node is drawn filled once it's digested — you've read it and written at least one claim off it — and as a hollow outline while it's only imported. So a board shows what you've actually taken in, not how big your library is. There's no percentage meter; a dense patch of filled nodes is the score. (The library view below spends fill on cluster identity instead, so it doesn't make this distinction.)

Les arêtes sont des relations typées

A link between two nodes carries a relation — the half that holds your thinking (supports, refutes, extends, …). Those are yours. The system also draws its own edges — cites, similar, anchor, link (citations, cosine similarity) — in a distinct, quieter style; they're the system's voice, never mixed up with yours. An edge is true everywhere: draw it on one board and it holds across the whole graph.

Lire la vue bibliothèque

The library view shows every paper and claim at once and lays them out by force, so position is not yours to set — it falls out of the edges. Four things carry meaning there:

Colour is a cluster. KL groups nodes into communities from the citations and semantic edges between them, and tints each one. The clusters aren't a folder structure you maintain — they're re-derived from the graph every time you open it, so they shift as you read and link. Hover a tinted region and it names itself with the words and phrases over-represented inside it compared with the rest of your library. Phrases matter: a shelf of papers on experimental design is badly described by "experimental" and "design" sitting apart, so a phrase and its own fragments never both take a slot. Comparing against the rest of your library is also why a cluster's label usually lands more specific than the field it belongs to. Very small groups, and anything past the eighth cluster, stay neutral grey rather than borrowing a colour that already means something else.

Size is connectedness. A node grows with the number of citations and links touching it, in either direction. Semantic neighbours barely count toward it — every node has roughly the same number of those by construction, so sizing on them would make everything the same size and tell you nothing. A big node is one your library keeps coming back to, either because much of it leans on that paper or because that paper reaches into much of it.

Shape is kind — a disc is a paper, a diamond is a claim — because colour is already spoken for.

Edge weight is confidence. Citations are drawn in the system's orange; semantic neighbours fade in and out with how strong the match actually is, relative to the range your own library exhibits.

Voisins mutuels ou tous les voisins

Semantic edges come from the top few nearest neighbours of each node, and that relationship isn't symmetric: a broad survey lands in many nodes' shortlists while its own shortlist holds only its closest few. By default the view keeps only mutual neighbours — pairs where each ranks the other — because the one-sided ones wire a popular paper into every neighbourhood at once and melt the clusters together. all neighbours restores them if you want to see everything the embedder proposed; citations only drops semantic edges entirely, leaving just what authors and you actually asserted. The similar-K slider sets how many neighbours each node nominates before that filter runs.

Les tableaux sont des dispositions, pas des graphes

A board is a named, hand-laid-out window onto a corner of the one graph — "Chapter 3's literature", "the attention lineage". It stores only positions, never edges, so boards never fork the truth; they're just places to arrange a slice of it.

La disposition porte le sens

Where you put a node is information. When you export a board to a draft, KL reads it the way you read a page — left to right, top to bottom — and that becomes the order of your sections. Nodes you've linked nest under what they support or extend; nodes you've simply placed in a row become a sequence. So arranging the board is already outlining the argument — there's no separate ranking step. → outline follows your links; → outline (story) follows pure board order.

Un coin du graphe, en détail

You read Bahdanau and Vaswani and wrote a claim off each (filled — digested), so you could honestly link them. Sutskever is cited by Vaswani but you haven't read it yet, so it stays a hollow outline. Citations are the system's voice (orange); the extends edge between your two claims is yours.