Classify entities and build collections

Write for a few weeks, then open your knowledge graph. You'll see hundreds of dots: the people you work with, the tools you're weighing, the ideas you keep coming back to. A raw graph won't tell you which dot is which. Two tools fix that. Kinds say what each entity *is*. Collections gather the entities you care about into living sets that keep themselves up to date.

Neither one asks you to file anything by hand. The extractor sorts entities as it reads your notes. A collection is a rule you write once, like "People connected to Acme", and it keeps answering itself as your workspace grows.

The 11 core kinds and custom refinements

The 11 built-in kinds provide a shared base for graph colors, tables, and filters. In Ontology, you can define custom kinds with a core parent. The extractor can select these refinements; if the evidence does not justify a kind, it leaves the entity untyped.

Each kind gets a fixed color, and it's the same color everywhere: the dot on a graph node, the row in the entity table, the badge on a page. An entity with no kind shows up as "Type unknown", which just means no type was found yet.

Votes, ties, and two badges

A kind is a per-note observation. Each note that mentions an entity casts its own vote. The vote is recorded in that note's own slice of the graph, so it remembers which note cast it. The type you see is the modal vote: the most common one across all your notes, plus any merged synonyms.

# What the meeting note saw
(Apple) --is a--> (Organization)

# What the gear note saw
(Apple) --is a--> (Product)

When the votes tie, the entity shows two badges, on purpose. Language really does use one name for both a company and its products; linguists call this metonymy. Showing both keeps the information visible.

• Your manual override always wins. Set it from the type pill on an entity page (an untyped entity shows a dashed "Set type…" pill), or by clicking a chip in the "Entity types" panel.
• Extracted from the note comes second: whatever the extractor observed as it read.
• AI backfill ("Fill in missing types") comes last, and a backfill vote never overrides what a note actually said.

Overrides live outside those per-note slices of the graph. That's why re-extracting your workspace can never wipe out a type you set by hand. To undo one, pick "Clear type override" from the same menu.

The Entity types panel and "Fill in missing types"

You'll find the whole inventory in the knowledge-graph view's left rail, under "Entity types". A summary line tells you how many entities have a known type and how many don't.

Below it, one card per kind shows a color dot, a count, a 12-chip preview with "Show all", and an "Open as table" button. A dashed "Type unknown" card holds the rest. Click any entity chip to set its type by hand; the menu lists all 11 kinds plus "Clear type override".

Collections that update from saved rules

A collection is a saved set of entities, stored as a recipe of steps that re-runs against the current graph every time it's drawn. "People connected to Acme" is still right next month, because next month it gets computed all over again from next month's notes.

• Start from a kind (all your People), from entities you pick by hand, or from a mention count.
• "Only ones connected to…" keeps just the members linked to an entity you name.
• "Hop to a connected type…" or "All connections" walks the whole set along its edges to its neighbors.
• "Filter by mentions" keeps only entities mentioned at least N times.

You build and save collections from the round Collections button floating on the graph canvas (there's a Collections tab in the tuning panel too). Once saved, a collection shows up as a source everywhere. The entity table's source switcher lists it with a live member count, and saved rows offer "Open as table".

Deleting one removes only the recipe. The confirm dialog says the same thing before you commit. What goes away is the saved steps; your notes and entities stay untouched.

Build one, step by step

Key terms

Entity kind

A classification of an entity, recorded as a per-note observation. Eleven core kinds provide the base; custom kinds add domain-specific refinements.

Type unknown

No kind was found: the extractor said "other" instead of guessing.

Modal vote

The kind you see: the most common per-note observation across all notes and merged synonyms.

Manual override

A type you set by hand. It outranks every other vote and survives re-extraction.

AI backfill

"Fill in missing types": a quick classification from the name and description only, and the lowest-priority vote.

Metonymy

One name standing in for related things (a company and its product). A tied vote shows two badges.

Collection

A living set of entities saved as a recipe of steps, re-run against the current graph every time it's drawn.

Every entity now carries what it is, and collections keep the sets you care about fresh. Next, in domain vocabulary, you teach the AI the relationship words of your own field.