Domain vocabulary: teach the AI your field

Lesson 1 of 310 min

What you'll learn

you can explain why an open verb vocabulary fragments, and how canonical predicates pull the synonyms into one edge
you can fill in the "Domain vocabulary" panel: domain, reading persona, and connection types with variants
you can auto-tune with "Teach the AI your field" and adjust what it learned before saving
you can say why saved vocabulary only shapes notes from now on, and re-extract old ones to catch up

Monday's note says Acme built the Orbit app; Wednesday's, half in Korean, says Acme가 Orbit 앱을 개발하다; Friday's says Acme developed it. You know it's the same fact, but the extractor records each verb exactly as written. Your graph gets three thin parallel links where one strong one belongs.

The fix is the "Domain vocabulary" panel. It is the first of the five steps "Tune up AI search" runs, which is why the panel reads "Stage 1 / 5 · Foundation". You teach the extractor your field once: what the domain is, how to read your notes, and which connections matter under which canonical names. From then on, every note you save is extracted with that vocabulary in mind.

Why an open vocabulary fragments

Recall from the ontology lesson: entity kinds are a closed list of 11, but relations are deliberately open. When the AI reads a fact, it records the connection under whatever verb the sentence used, like built, developed, or 개발하다.

That openness lets the graph describe any field without waiting for someone to approve a schema. The catch is drift: synonyms quietly split into separate families of connections.

Three notes, one relationship, three edges

Acme  --built------->  Orbit
Acme  --developed--->  Orbit
Acme  --개발하다------>  Orbit

Same real-world fact, three parallel edge families.

The panel: three fields

Open the knowledge graph ("Open knowledge graph" in the titlebar) and pick "Domain vocabulary" in the left rail. The panel's header reads "Teach the AI your field". It edits exactly three things:

Field	UI label	What you write
Domain	"Domain" (placeholder "e.g. space systems")	One phrase naming your field.
Reading persona	"How the AI should read your notes" (placeholder "e.g. Read these like an aerospace systems analyst…")	The lens the extractor reads with: what a careful expert in your field would notice.
Connection types	"Important kinds of connections"	A list of canonical predicates. Each has a name (e.g. "reports to"), a one-line description, and "other ways to say it (comma-separated)" for the variants. "Add a connection type" adds a row; the X removes one.

domain: A one-phrase name for your field, added to every future extraction.
reading persona: How you want the extractor to read your notes: the expert lens it puts on.
canonical predicate: The one preferred name for a kind of connection, e.g. "develops".
variant: Another way of saying a canonical predicate ("made", "built", "개발하다") that folds into it.
write-time folding: Variant verbs are swapped for their canonical predicate the moment a fact is stored, before it lands in the graph.

Auto-tune: let it learn from your notes

You don't have to fill any of this in by hand. The footer's "Teach the AI your field" button reads the notes you already have and drafts all three fields. Treat the result as a draft: rename a predicate, drop the ones you don't care about, add the variants you actually write.

Tune the vocabulary in four steps

Open the panel
Knowledge graph view → left rail → "Domain vocabulary".
Press "Teach the AI your field"
Wait while "Learning your field…" runs. It reads your existing notes to draft the domain, persona, and connection types.
Adjust the draft
Edit the "Domain" and persona text, then tidy up "Important kinds of connections": the names, the one-line descriptions, and the comma-separated variants. Use "Add a connection type" for anything it missed.
Save
Press "Save" and look for "Saved". Note the hint beside it: "Applies to notes from now on".

You don't even have to remember this panel exists. As your workspace grows, the activity center offers "Tune domain vocabulary?" and gives you three buttons: Tune, Later, and Don't ask again.

The Lifecycle panel's one-click "Tune up AI search" also starts with Domain vocabulary. Re-extraction, the synonym check, the relationship check, and the topic map follow.

Write-time folding: many verbs, one edge

Your vocabulary rides along with every future extraction, so the extractor reads with your persona and prefers your predicate names. When it writes a fact whose verb matches a variant, that verb folds into the canonical predicate at write time. The graph stores one edge family instead of three.

After teaching: "built" / "developed" / "개발하다" → develops

Connection type: develops
  description: who built or created what
  variants:    built, developed, made, 개발하다

New note:  "Acme built the Orbit app."
Stored as: Acme --develops--> Orbit

Quick check

You add "built" as a variant of "develops" and press "Save". A note from last month still shows a "built" connection on its entity page. Why?

Making it retroactive: re-extraction

To bring old notes up to the new vocabulary, re-extract them. There are two routes. The rail's "Extraction" section has a "Re-extract everything" button that re-reads every note with your saved vocabulary in mind.

The other route is the Lifecycle panel's "Tune up AI search". Tick "Re-read every note (slower, but more thorough)" and re-extraction rides along with the full tune-up pass.

Quick check

Which three things does the "Domain vocabulary" panel actually edit?

Wrap-up

Relations are an open verb vocabulary. That's expressive, but synonyms fragment into parallel edges.
"Domain vocabulary" teaches the extractor three things: a domain, a reading persona, and canonical predicates with variants.
"Teach the AI your field" drafts all three from your existing notes; you adjust and press "Save".
Variants fold into canonical predicates at write time, so this covers future notes only. "Re-extract everything" brings old notes along.

Now the graph speaks your verbs. Next come statements: your own decisions, captured in the sentences you actually wrote.

← Previous lesson