Document saves and graph extraction

It's 4:58 pm. You finish a project meeting note (who owns what, which deadline slipped) and hit save. A few seconds later, Consilience knows that Mina owns the launch checklist and that the deadline moved to March.

You didn't click an import button, and nothing left your machine except one AI call. All you did was save a markdown file.

Search finds the sentences you already wrote; this lesson opens up how the app learns what they mean. You'll see what background extraction does and how it decides whether a save is worth re-reading. Then you'll trace any fact back to the sentence that states it.

Extraction after a document save

With "Automatic knowledge-graph extraction" on (off by default on desktop; enable it in Settings), saving a markdown note can kick off a background read. The app never edits your file: it reads a copy, runs it through the extraction model, and records what it learned in the knowledge graph.

What it records is the entities the note mentions (people, projects, places, ideas) and the named relationships between them. If "Capture your decisions and questions" is on, it also records your own verbatim statements.

1. You save the note. The app checks the change gate first (next section). If the body hasn't changed, there's nothing to re-read.
2. A changed note joins a background queue and waits its turn.
3. The AI reads the note and records the entities, relationships, and statements it finds.
4. The results go into that note's own named graph in the local store, replacing only what this note said before.
5. The UI catches up live: the file-tree badge clears, the "Key facts" count updates, and the note's connection graph refreshes.

Document changes and re-extraction conditions

Every note carries a quiet fingerprint of its body text, and that fingerprint is the gatekeeper. When you save, the app compares the old fingerprint against the new one. If they match, there's nothing new to read, so the AI is never called. Change a single character and the fingerprint changes too, so the note re-extracts and its facts are rewritten from scratch.

• Even a one-line note gets extracted. A single character of body is enough. Empty notes, or ones with only whitespace, are left alone.
• "Re-extract everything" (in the graph view's "Extraction" panel, which a later lesson tours) and Settings' "Rebuild knowledge graph" clear the gate on purpose, so every note re-queues under the latest logic.

Reading the file-tree badges

The file tree in the Workspace sidebar answers one question at a glance: has the AI read this yet? A colored dot sits at the right edge of each file row (folders and excluded files never get one). Hover or click the dot and a card pops up that names the state and offers a one-click fix:

Click the button and the file joins the queue. The badge flips to "Adding…" and a toast confirms "Adding to AI search…". A row with no badge at all is the state you want: the file is fully current, and what the AI knows about it matches what's on disk.

Key facts: audit what a note asserted

Open any markdown note and look at the document sidebar. A "Key facts" button carries a live count of the facts pulled from this note. Click it and the "What this note says" dialog opens. This is your per-note audit view. Before anything is extracted, it just says "No key facts yet".

• "Key facts" tab: one readable card per entity the note mentions. You get the name in bold, a kind chip (like "Person"), an "Aliases" line, a one-line description, then relation rows linking it to other entities.
• "Connections" tab: this note's neighborhood, laid out as lists. Backlinks, unlinked mentions, notes related by meaning, and knowledge-graph neighbors. Click any of them to open that note.
• "Connection graph" (right side): "The people, projects and ideas this note connects to." It's everything one step away from this note, drawn with the same 3D view as the full graph you'll tour later. Click a node to isolate its neighborhood, click it again to exit.

Every relation row can prove itself. Hover one and Consilience shows the exact sentence in your note that states the fact, with its line number and the heading it sits under. It's always checked against your note's current text, so it stays accurate. Edit the note and, instead of quoting something that's no longer there, it says "Couldn't find a sentence stating this. The note may have changed."

The dialog is built to be read by a person. Entities that show up name-only (no facts, aliases, type, or description) are set aside as noise, so what's left is only what the note actually asserted.

Your markdown is never touched

All of this happens in a separate store beside your files, one per workspace, kept on your device. Extraction reads a copy of your note and writes only to that store.

Your notes stay byte-for-byte what you wrote. The graph is just a rebuildable index, and Settings' "Rebuild knowledge graph" regenerates the whole thing from your files whenever you want.

background extraction

The on-save pipeline that reads a changed note and writes entities, relationships and statements into the graph.

change gate

A quiet fingerprint of a note's body text. If it hasn't changed, the note is never re-read.

named graph

The per-note compartment in the store that holds exactly what that note asserted. It is the provenance record.

lifecycle badge

The colored dot on a file row telling you whether the AI's knowledge of that file is current.

source sentence

The exact sentence in your note that states an extracted fact, shown with its line number and heading.

Recap

• Saving a changed note quietly re-extracts it. An unchanged save is never re-read.
• The file-tree dot is your dashboard: red = not in, orange = behind, nothing = current.
• "Key facts" is the per-note audit view, and each fact links to the exact sentence that states it.
• Extraction writes only to the separate store. The graph is a rebuildable index.

That's one note, arriving quietly. Bringing in the hundreds of files you already have, all at once, is what the Import lesson and its Extract sources wizard are for.