Digest

Digest

You have material to add to the knowledge graph. This skill walks you through a complete ingestion: extract → create → link → annotate → verify.

Workflow

1. Check what already exists

search(kind="entity", query="<key terms from the material>")

Never create duplicates. If the entity exists, skip to linking or enriching it.

2. Create entities

For each nameable concept, paper, person, project, dataset, or org in the material:

create(kind="entity", entity_kind="<kind>", name="<short canonical name>",
  description="<1-2 sentence summary>",
  properties={"domain": "...", "type": "...", "year": "..."})

6 entity kinds (closed — pick the best fit, don't invent):

Kind	Use for
concept	Algorithms, techniques, architectures, models, research gaps
document	Papers, preprints, reports, blog posts
dataset	Benchmarks, corpora, evaluation sets
project	Codebases, libraries, tools, frameworks
person	Researchers, engineers, authors
org	Labs, companies, institutions

Naming: short canonical name people actually say. LoRA not Low-Rank Adaptation of Large Language Models. Full titles go in properties.

3. Link entities

For each relationship you identified in the material:

link(source_id="<from>", target_id="<to>", relation="<relation>", weight=<0.4-1.0>)

13 relations (closed — map to these, don't invent):

Category	Relation	Direction	When
Structure	contains	parent → child	System has component
Structure	part_of	child → parent	Inverse of contains
Structure	instance_of	specific → general	X is a case of Y
Derivation	extends	child → parent	Builds on, generalizes
Derivation	variant_of	variant → original	Modified version
Derivation	introduced_by	concept → paper/person	First described in
Derivation	supersedes	new → old	Replaces entirely
Dependency	depends_on	consumer → dep	Hard requirement
Dependency	enables	prerequisite → outcome	Makes possible
Implementation	implements	code → concept	Code realizes algorithm
Lateral	competes_with	A ↔ B	Alternative approaches
Lateral	composed_with	A ↔ B	Used together
Annotation	annotates	note → any substrate	Note observes/comments on

Direction matters. introduced_by goes FROM the concept TO the paper (the concept was introduced by the paper). If you get direction wrong, the traversal breaks.

Weight: 1.0 = definitional, 0.7-0.9 = strong, 0.4-0.6 = plausible.

4. Create notes (observations, insights, decisions)

For anything worth recording that isn't a nameable entity:

create(kind="note", note_kind="<kind>", content="<the observation>",
  salience=<0.0-1.0>, annotates=["<entity-uuid>"])

5 note kinds: observation (I noticed), insight (I concluded), question (I wonder), decision (I chose), reference (I read).

Always use annotates to attach notes to the entities they're about.

5. Verify density

After ingestion, check each new entity:

neighbors(node_id="<entity-id>", direction="both")

Minimum edges: concepts ≥ 4, projects ≥ 3, documents ≥ 2. If below target, add edges — every concept should have at least instance_of or extends (parent), introduced_by (if a paper exists), and competes_with (if alternatives exist).

6. Report

Summarize: what entities were created, what edges link them, what notes captured, and what gaps remain (questions filed as question notes).

Stop condition

Material exhausted. Every entity above minimum density. No orphans (0-edge nodes). Gaps identified as question notes for follow-up.

Error handling

If a tool returns an error, read the message — it lists valid values. Common cases:

• Invalid entity_kind or note_kind → the error says which values are valid
• Invalid relation → use only the 13 above
• ID not found → check the UUID; use search to find the correct one