Salesforce org-intelligence glossary
Short, factual definitions for terms used across sf-intelligence docs. Each entry is written to be citable by search engines and AI assistants.
MCP server (Model Context Protocol)
A standardized way for AI assistants (Claude, Cursor, Codex) to call external tools. A Salesforce MCP server exposes org-aware capabilities - schema lookup, dependency analysis, permission tracing - that the model invokes during conversation. sf-intelligence is a read-only Salesforce MCP server focused on org understanding, not DevOps. See the MCP overview.
Metadata vault (org-kb/)
The local directory created by sfi refresh. Contains Markdown-rendered component docs, a DuckDB dependency graph, manifest files, and optional overlays (annotations, git history). All offline answers are served from here - no network calls during conversation.
Refresh
A read-only metadata pull via sf project retrieve followed by local indexing. Re-run after org changes to keep answers current. Large orgs can use sfi refresh --staged for a servable vault in minutes. See staged refresh.
Intelligence planes
Offline - answers from the vault (default). Live - opt-in, read-only SOQL for record counts and samples. Hybrid - fuses vault structure with live magnitude, disclosing both provenances.
Provenance
Where an answer's data came from: offline_snapshot, live_org, hybrid, or annotation (human-curated overlay). Every sf-intelligence answer carries a provenance stamp.
Confidence tier (edge confidence)
How a single dependency edge was derived: declared (Salesforce metadata), parsed (AST/XML analysis), or heuristic (regex - verify before acting). This is the edge axis - one relationship's evidence. An interpreted claim carries a separate claim confidence; don't conflate the two.
Completeness
Whether all needed metadata families were retrieved: complete, partial (some types pending or not checked), or unknown.
Reasoning / concept model
A curated, org-independent library of structural patterns - 143 concepts / 195 rules of general Salesforce truth (master-detail cascade, junction objects, read-only formula/roll-up fields, stacked record-triggered flows, OWD sharing posture, and more) that carry no org data. sfi.interpret joins this concept model against your grounded org graph and runs it as a deterministic rule engine over an offline graph slice - no LLM inference, no live call - to turn structure into cited claims. The full path is resolve → interpret → synthesize_answer. It matches known patterns; it does not "understand" your org or hunt for vulnerabilities.
Interpretation (cited claim)
A single grounded claim emitted by sfi.interpret when a concept rule fires on a component - e.g. "deleting this master-detail parent deletes its child records," or "this roll-up field is read-only." Each claim carries the component ids it is groundedIn as citations: no citation, no claim. When no rule fires, that is stated plainly and is not a claim that nothing depends on the component.
Claim confidence
The confidence on an interpreted claim - a second axis distinct from edge confidence. A claim is held at the weakest of the firing rule's ceiling and the grounding edges it cites, so it is never asserted above the evidence beneath it. Static-shape claims describe structure, not live runtime behaviour.
Applied scope / scope refusal
Scope-aware tools accept a natural selector (a canonical componentId, a bare apiName/objectApiName/profileApiName, a nameContains substring, a namespacePrefix) and echo the resolved scope back as appliedScope. If selectors disagree or none resolves, the tool refuses with a named invalid-query rather than silently answering org-wide against the wrong target; omitting scope returns the full, byte-identical default. See configuration.
Impact analysis / blast radius
Tracing components that depend on a field, object, Flow, or Apex class before you change or delete it. Core use cases - see Salesforce dependency analysis and impact analysis.
Static analysis
Analysis of retrieved source and metadata without running code. Dynamic SOQL and reflective Apex are invisible. "No references" means no static evidence, not "unused."
Canonical component ID
A stable identifier like CustomField:Account.Industry__c or ApexClass:PaymentProcessor. Every component sf-intelligence names is cited with one.
Semantic router
The offline semantic layer that surfaces a meaning-ranked shortlist of the right tools (and intelligence plane) for each plain-language question; your AI host picks which to run. Logs unrouted questions locally for product improvement.
Org card
A compact JSON summary of org scale and coverage rendered at every refresh. Served via sfi.org_card for fast orientation.
Annotations overlay
Human- or AI-proposed metadata stored in meta/annotations.jsonl - owner, status, glossary synonyms - that survives refreshes. See configuration.
Read-only guarantee
sf-intelligence has no write, deploy, or delete path to Salesforce. Only metadata retrieve touches the org. See FAQ and trust.