Glossary

A quick reference for the terms used across the docs. (Sample page — demonstrates a granular, searchable reference doc.)

Core

RelayEvent : The canonical event envelope on the spine — the wire contract. ULID eventId is the platform-wide idempotency key.

Agent : The persistent per-(tenant, store) LLM reasoner. Phase-1 variant Agent.StoreAgent.single@1.0.

Suggestion : The agent’s unsolicited, quantified, routed recommendation.

Conversation / turn : The agent’s solicited, multi-turn answer to a user.

Decision layer

InterestingnessDetector : The universal, learned attention gate — scores each event for “worth the agent’s reasoning.” No per-tenant config.

SafetyNet : The ~10 hardcoded critical conditions that always escalate — the floor the platform never misses.

LlmCallRecord : The immutable audit record of one LLM call (full prompt + response + composition refs). The backtest and prompt-evolution substrate.

Knowledge & context

VerticalPack / TenantContext / StoreContext : The three layers of structured knowledge that describe a business (not its rules) and compose into the agent’s prompt.

Knowledge scopes : Retrieval is scoped platform → enterprise → store; the most specific layer wins.

Tenancy

TenantResolver : The single multi-tenant safety control. Resolves the tenant, injects a scoped context, fails closed (401). App code never reads tenantId directly.