Connect

Connect

You just encountered something new and want to integrate it into your existing graph. This skill finds what it relates to and wires the edges.

Workflow

1. Check if it already exists

search(kind="entity", query="<the new thing>")

If it exists: skip to step 3 (enriching edges). If it's a close match but different (e.g., a variant): create it and link via variant_of or extends.

2. Create the entity

create(kind="entity", entity_kind="<kind>", name="<short canonical name>",
  description="<what it is>", properties={...})

Pick from 6 kinds: concept, document, dataset, project, person, org.

3. Find what it connects to

Search broadly for related entities:

search(kind="entity", query="<related terms, parent concept, enabling technique>")

Think about these relationship dimensions:

• What is it a kind of? → instance_of (specific → general)
• What does it build on? → extends (child → parent)
• Who introduced it? → introduced_by (concept → paper/person)
• What does it need? → depends_on (consumer → dependency)
• What does it compete with? → competes_with (A ↔ B)
• What does it enable? → look for entities that depends_on prerequisites this fulfills
• Where is it implemented? → implements (code → concept)
• What is it used with? → composed_with (A ↔ B)

4. Create the edges

For each identified relationship:

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

Direction rules (the most common mistakes):

Relation	Points FROM → TO	Mnemonic
introduced_by	concept → paper	"LoRA was introduced_by Hu 2021"
extends	child → parent	"QLoRA extends LoRA"
instance_of	specific → general	"GQA is an instance_of grouped attention"
implements	code → concept	"lattice-inference implements GQA"
depends_on	consumer → dependency	"quantization depends_on calibration data"

If the relationship doesn't fit any of the 13 relations, it's probably a property on the entity (e.g., "published in 2021" → properties.year: "2021", not an edge).

5. Verify density

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

Targets: concepts ≥ 4 edges, projects ≥ 3, documents ≥ 2. If below, actively seek more connections — especially instance_of/extends (every concept has a parent) and introduced_by (most concepts have a source).

6. Report

State what was connected: the new entity, each edge created (with direction and rationale), and final edge count. Flag if density target wasn't met and why.

Stop condition

Entity is at or above minimum density. Each edge has clear rationale. If you can't reach the density target, note what's missing (e.g., "no source paper found for this concept — needs research").

What is NOT a relation

These are properties, not edges:

• "published in 2021" → properties.year
• "uses Python" → properties.language
• "has 7B parameters" → properties.params
• "achieves 95% accuracy on X" → properties.benchmark_results
• "is popular" / "is important" → not an edge; maybe salience on a note about it