Core concepts

A short vocabulary that the rest of these docs assumes. See the full Glossary for more.

Context graph

The context graph is the shared, graph-based semantic memory your agents reason over — the canonical model of your enterprise’s entities and relationships. As data arrives through connectors, the platform self-organizes the graph: it discovers entities, relationships, events, and facts and weaves them in.

Retrieval works by graph traversal: rather than flattening your corpus into nearest-neighbor text search, Fabric starts from the entities relevant to a query and follows the graph’s real relationships to assemble the connected context that actually matters. The graph is never exposed directly — every query is scope-pinned to your project, read-only, and bounded.

Ontology

The ontology is the self-organizing schema of entity and relationship types discovered from your data. In Simple Mode it is created and extended automatically as data arrives. In Advanced Mode you define it explicitly through a proposal → commit flow, with human review of low-confidence mappings.

Scope: organization → department → project

Every object in Fabric carries a scope in the hierarchy organization → department → project. The active scope flows on every request, and the server rejects any request whose scope the caller cannot access. A department can tighten but never loosen the grants it inherits, giving clean business-unit isolation. When you connect an agent, its credentials are pinned to a single workspace — which maps to a project.

Governed action

When an agent proposes an action, it runs through a fixed governed sequence before it ever touches a production system:

1. Policy check — policy-as-code returns allow, deny, or allow-with-gate; the decision is recorded with its reasoning.
2. Simulation / dry-run — the action runs against shadow state only, producing a before/after diff with zero impact on production.
3. Approval gate — high-risk actions pause for a single-use, time-boxed human approval.
4. Execute — only then does the action run, with status streamed back in real time.

Trust surfaces: trace, audit, lineage

Every run is a distributed trace — an ordered waterfall of steps with timing, token, and cost detail. Every consequential event is written to a tamper-evident, hash-chained audit log (never sampled). Lineage answers “which data influenced this action”. These surfaces are produced as by-products of the same execution pipeline, so they are complete and contemporaneous rather than reconstructed after the fact.

MCP (Model Context Protocol)

MCP is the open standard by which agents discover and call tools. Fabric is both:

• an MCP server — your agent connects in and calls Fabric’s context tools;
• an MCP / A2A client — Fabric reaches out to external tools and peer agents.

See The Fabric MCP tools for the tools Fabric exposes inbound.

Simple and Advanced mode

Every project runs in one of two modes, set per project:

Mode	UI label	Ontology	Data writes	Approvals
Auto	Simple	Self-organizing — created and extended automatically as data arrives	save-context writes nodes and relationships directly	None (frictionless)
Manual	Advanced	Explicit proposal → commit	Not allowed via save-context	Commit may require human approval

In both modes every call is scope-pinned, policy-checked, and audited — Simple mode removes approval friction, not governance. See Simple vs Advanced mode.