Cortex models your business as typed objects, instances, and relationships — then governs access down to the individual property and action. An operator can read a Claim but never its policyholder's ssn. Palantir-class semantics, no lock-in.
One cortex schema · multi-tenant isolation · deny-by-default
Embedding-only retrieval can quote a PDF, but it has no concept of a Claim, who filed it, or which Payout it triggers — and no way to hide one property from one role. Cortex gives agents a structured, governed object model so every answer and every action is grounded in real business entities, not loose passages.
Register object types with properties and allowed actions; create object instances; connect them with typed relationships. The Object Explorer computes the live neighborhood subgraph on demand.
Register object types — key, name, properties (jsonb), and allowedActions — as the governed schema. An object's type must exist before any instance can be created.
Upsert object instances keyed by (typeKey, externalId), then connect them with typed relationships like files, triggers, belongs_to. Idempotent — safe to re-sync from live platform resources.
GET /objects/:id/explore returns the neighborhood subgraph — typed nodes and relations — so agents reason over real entities, and relationship-aware retrieval beats document-only RAG.
Grants are rows of (role, type, property, access ∈ none/read/write). A type with no grants is ungoverned — full access. The moment any grant exists, that type flips to deny-by-default. A property grant overrides the type default; otherwise the property inherits it. Across a caller's roles, the most-permissive access wins.
Editing the object model is never silently final. Capture the live schema as a draft, publish it (archiving the prior version), or roll the live schema back to any prior version. Rollback restores from a snapshot and records a new published version — the target version is never mutated.
Document-only RAG retrieves passages. A governed ontology gives agents typed entities, relationships, per-property access, and ontology-scoped actions — the foundation Palantir built Foundry on, kept inside Cortex's Trust Layer and portable across tenants and air-gapped environments.
Ontology-scoped actions run dry-run → propose → approve → execute → compensate against typed objects.
Author allow/deny/require_approval rules that reference object types and properties.
Every object read, write, and version change lands in the hash-chained, provable audit.
Live ontology counts — types, objects, links, grants, current version — surface as KPIs.
Deny-by-default object and property permissions, redaction on read, and a versioned, audited schema — aligned with the data-governance controls your reviewers already expect.
Give every agent a governed object model — typed, versioned, and permissioned down to the property and action.