Graph

Obsidian’s own graph shows you the links you wrote. The Smart Graph shows you the structure you didn’t: the topics your vault has organized itself into, and the notes that are about the same thing without ever linking to each other.

Open it with Smart Second Brain: Open smart graph.

What’s drawn

Authored links. Your wikilinks, optionally with arrowheads showing direction (Directed links).

Tags. With Tags turned on under Scope, each tag is a node linked to every note that carries it, the way Obsidian’s own graph draws them with its Tags filter on. Tag edges are undirected and always drawn while tags are shown, whatever the link toggles say. Tags also shape the topics; see below.

Inferred links. When an embedding model is configured, the graph adds edges between notes that are semantically similar but not linked. Two controls govern them:

• Neighbors (k): how many inferred neighbours each note may contribute. The default is 5. The union is denser than it sounds, because an edge survives if either endpoint ranks the other.
• Similarity threshold: the minimum cosine similarity for an inferred edge. The default is 0.55. Below roughly that, pairs are noise rather than genuine topical overlap.

The threshold matters more than k for keeping topics clean.

Inferred edges are drawn dashed so they read as distinct from authored links up close, but at overview zoom the dashes sit only a few pixels apart and the two kinds blur together. Highlight inferred links (under Display, off by default) draws them in the accent colour instead, so the inferred layer stands out at any zoom. It only changes colour — width and dashing are unchanged, and it needs Inferred links turned on to have anything to colour.

Topics

The graph partitions itself into topics using Leiden community detection, and renders them as tinted regions with name pills.

Granularity is the control you’ll actually use: one slider that moves between few broad topics and many narrow ones. Underneath it is the Leiden Resolution (γ): higher means more, smaller topics.

Because Leiden returns a flat partition, the hierarchy is built by running it twice, once coarse and once fine, then rolling each fine topic up into whichever coarse topic holds most of its notes. That’s what lets you ask “what is this vault about?” at more than one altitude and get consistent answers: every note rolls up somewhere.

Tags shape topics

For a vault organised by tags rather than links, tags are the authored structure, so when tags are shown their edges join topic detection alongside your wikilinks, in both fused and link-only mode. A tag-organised vault no longer reads as “nothing is linked”.

Each tag’s pull is damped by how many notes carry it: a topical tag on a dozen notes binds them firmly, while a status tag on hundreds (#todo, #draft) can’t drag unrelated notes into one topic. A tag on a single note counts the same as one wikilink.

Tags are never topic members. They don’t count toward a topic’s size, carry no topic colour of their own, and don’t fold into a collapsed bubble. A tag that lives in a topic (at least half of the notes carrying it are inside) is drawn within that topic’s region, pulled toward its centre, and can give the topic its name.

Related controls:

• Link-only topics: detect topics from authored links alone (your wikilinks, plus tags when they’re shown), ignoring inferred edges.
• Seed: the PRNG seed. The same seed on the same graph gives the same topics, so a layout you liked is reproducible.

A group needs at least two notes to count as a topic. A single note is not a topic. It stays on the graph without a region.

Topic labels

Topics are named after their best-connected note’s filename by default. With a Topic naming model set under Settings → Graph, the Name topics with AI button in the Topics panel sends each topic’s note titles to that model and names the topic from them. Names are cached by topic membership, so an unchanged topic is never named twice. Naming runs in the background and can be cancelled, and if the graph changes underneath it, the in-flight pass is aborted. Name topics automatically (off by default) runs it whenever the topics change instead of only on request.

Highlighting

Two structural highlights are the reason to keep this view open:

Bridges. Notes whose neighbours are mostly in other topics. These are where your thinking connects, and they are usually the notes worth developing. The Bridge threshold sets the minimum fraction of foreign-topic neighbours required to qualify.

Isolated / unlinked. Notes with no authored links. Orphans you meant to connect and forgot.

Layout

The layout is a force simulation with five exposed forces:

Control	Effect
Link distance	Target distance between connected nodes
Repulsion	How strongly nodes push each other apart
Center force	How strongly the graph is pulled toward the centre
Link strength	How strongly edges pull connected nodes together
Cluster cohesion	How strongly nodes are pulled toward their topic centre

Rendering goes through PixiJS, and projection, clustering, and the semantic neighbour scan run in Web Workers so dragging stays smooth on a large vault. Past a threshold the neighbour scan switches from an exact pairwise comparison to an HNSW-accelerated approximate one, where the quadratic scan stops being viable.

Scope

Markdown only excludes non-markdown files from the graph. As with search, the agent folder is always excluded.

Tags (off by default) draws each tag as a node linked to the notes that carry it, and lets shared tags help shape the topics. Tag nodes use the theme’s tag colour and size by how many notes carry them; the tooltip shows the count. Clicking a tag node, or Search notes with this tag in its context menu, opens Obsidian’s search on that tag, the same as clicking a tag in a note. Tags match case-insensitively with the first-seen casing as the label, and nested tags are their own nodes: #a/b is not a child of #a. Only notes that pass the other filters contribute, so a folder or extension filter carries over to which tags appear.

Interacting

Clicking a note opens it. Click a topic’s label to collapse just that group, and shift/⌘-click to select more than one segment. Right-click a note for Open file, Reveal in file explorer and Focus on this cluster. On mobile, long-press for the same menu, since touch has no right-click. A tap highlights the note and its neighbours the way hovering does on desktop.

Selecting

The lasso in the toolbar draws a freehand loop around whatever you want. On desktop you can also hold Shift and drag to do the same without switching the mode on. On a phone the toolbar button is the way in. Clicking a topic’s pill selects that whole topic.

Whatever is selected, a bar appears with what you can do to it: Zoom to selection (F), Collapse or Expand if the selection is whole topics (C), Immerse (I), Open in chat (A, which reveals the chat with those notes attached), and Open all in new tabs (O). Esc clears the selection. The shortcuts are single keys, and work while the graph has focus.

On mobile the same actions arrive as a bottom sheet instead of a bar, with each verb as a full-width row.

Immerse

Immerse rebuilds the graph from the selected notes alone, and re-runs topic detection over just that subset — so the topics you get are the divisions inside what you selected, which the full graph was too coarse to show. A topic that was one region becomes several. You can immerse again from there.

Granularity is remembered per graph, so dialling the topics finer inside an immersion describes that subset only. Leaving restores the setting the full graph had. The bar names what you’re in (“Immersed into Sleep & memory”) and carries an Exit back to the whole vault. Esc leaves too, once nothing is selected.

On mobile the selection sheet closes as soon as you immerse, so it never covers the graph you just moved into — the name of what you’re in moves to a banner top-left, and the exit becomes a highlighted button on the toolbar. Both stay put while you pan, select and immerse further.

Note

Inferred links, semantic topics, and generated topic labels all depend on configuration you may not have: inferred edges and semantic topics need an embedding model, and generated labels need a chat model. Without either, the graph still renders your authored links and detects topics from them alone.