SkyAIApp glossary

One full execution of a multi-step task — includes LLM calls, tool calls, state. Each run has its own trace_id and budget.

Per-request hard constraint (e.g. maxCostUsd). Router refuses candidates that exceed it; exceedance returns 422 router.budget_exceeded.

Embedding-similarity-based cache. Similar prompts count as a hit. Distinct from exact-match caches.

The eligible model set after the router filters by modality / context window / RBAC / budget — only candidates enter scoring.

Human-readable explanation of why the router chose what it chose. Lives at res.routing.decisionReason for audit + debug.

First layer in front of the router — Cloudflare edge does TLS termination, WAF, rate-limiting.

Workflow that scores model outputs. Used for A/B comparison, regression testing, guardrail reinforcement.

Ordered list of backup models tried when the primary fails. Each candidate re-runs budget validation.

xxhash64 of the normalized prompt. Used for idempotency keys and cache keys.

High-level intent: cost / quality / stability. Sets the primary axis the router optimizes against.

Around-request interceptor: PII detection/redaction, content moderation, prompt-injection defense. Has dedicated trace spans.

Client-generated unique request identifier. Same key within 24h returns the first response with no double-billing.

Per-request K/V tags (e.g. tenant, workflow). Indexed into traces and billing for filtering, analytics, quotas.

Logical partition of the cache. Same-namespace queries share; different namespaces isolate. Recommended split: tenant + workflow.

Named rule-set configured in the console — includes goal, strategy, fallback, budget, guardrails. Reference by policy_id.

Upstream model vendor (OpenAI, Anthropic, Google, DeepSeek, Mistral, Meta…). Router picks across providers.

Role-based access control. Governs who can create keys, edit policies, read traces. SAML/SCIM integrated.

The process of choosing primary + fallback models for a single request. Inspectable, traceable, replayable.

Isolated environment (api-sandbox.skyaiapp.com). Same schema, never billed, returns canned model output. Use in CI.

A node in the trace tree. Represents one operation (cache.lookup, execute.primary, http.openai…) with start time, duration, status.

Refinement under the Goal: balanced / cost-optimized / quality-first / latency-optimized. Sets weights for (cost, quality, latency).

Third-party services SkyAIApp uses (AWS, GCP, Cloudflare, etc.). Full list + 30-day notice window for additions.

External capability an agent can invoke — built-in (web_search, calculator, code_exec) + your custom tools registered via JSON Schema.

Complete decision record of one request. A span tree with routing decision, fallbacks, cache, guardrails, model HTTP calls.

Platform events pushed to your endpoint (router.fallback_triggered, etc.). HMAC signed + auto-retry + 24h replay.