Knowledge Modeling

Domain model, ontology validation, and knowledge-graph extraction — the typed-vocabulary family.

Domain modeling (active discipline)

Build and sharpen the domain model as you design — challenge terms, invent edge-case scenarios, write glossary + decisions the moment they crystallise. (Merely reading .devin/CONTEXT.md is a habit, not this skill — this is for changing the model.)

Structure: .devin/CONTEXT.md (glossary + bounded contexts) + .devin/adr/ (decisions). Formats: reference/CONTEXT-FORMAT.md, reference/ADR-FORMAT.md.

Cross-skills: research for authoritative term definitions; deep-mode for how code currently uses terms; ai-coding-dictionary for AI-coding jargon alignment.

Ontology validation (before side effects)

Pydantic at the door, ontology at the ledger.

Schema/type checks (Pydantic or equivalent) validate shape; ontology validates domain-level reasonableness. Check, before any write/insert/API call driven by tool output:

1. Entity type valid for this domain.
2. Status value allowed (pending/in_progress/completed…).
3. Relationship type makes sense (contains, depends_on…).
4. No duplicates — entity or relation already exists?
5. Disjointness — disjoint entities treated as same?

# Schema validates shape
class Ticket(BaseModel):
    status: Literal["pending", "in_progress", "completed"]

# Ontology validates domain
if ticket.status == "completed" and ticket.depends_on_open():
    raise ValueError("Cannot complete ticket with open dependencies")

Helper: python3 scripts/validate.py (this skill's scripts/). Anti-pattern: trusting tool output because the schema allows it.

Knowledge extraction (text → typed graph)

Deterministic, provider-neutral extraction of entities, relations, evidence, and provenance from Markdown/plain text into versioned JSON + relational Markdown under .devin/. No API key, no LLM — lexical baseline (not vector search).

• Every extracted fact points back to source file:line + quote.
• Incremental: merge new sources without duplicating entities or silently overwriting conflicts (conflicts reported, explicit action required).
• Baseline for evaluating Hyper-Extract/embeddings/MCP later.

Helper: python3 scripts/extract.py.