Production AI platform

Hybrid model cascade, RAG, MCP/API tools, caching, guardrails, events, and operations — hover any node for its failure mode or rationale
Live request Async processing Control & security State & caching Failure & degradation
Entry and control plane
Users & applicationsWeb, mobile, internal agents, API clients
API gatewayTLS, authentication, quotas, rate limits
Policy layerTenant, role, data class, budget, consent
Input guardrailsPrompt-injection screen, content safety, PII flags
AI orchestratorRouting, tool loop, workflow steps Stateless — state externalized
State & caching plane the latency and cost levers; every cache key carries tenant identity
Session & state storeHistory, agent checkpoints, durable workflow state
Semantic response cacheNormalized query → validated answer, TTL-bounded
Prompt prefix cacheSystem prompts, stable context at the model layer
Embedding & retrieval cacheHot query vectors and result sets
Orchestration execution paths selected automatically per request; any combination may run
Knowledge / RAG
Query plannerRewrite, decompose, apply access filters
Retrieval serviceHybrid search, rerank, deduplicate Live ACL re-check · keyword-only fallback
Context builderEvidence selection, token budget, citations
Actions / live data
Tool policy gatewaySchema validation, approvals, call limits
Direct APIs or MCP clientDelegated user and service identity Idempotency keys · saga compensation
Domain APIsFinal authorization and business rules
Model cascade
Route classifierPredicts tier before generation
Always-on SLMRoutine extraction, classification, answers
Warm self-hosted LLMComplex or private workloads
Hosted fallbackCapability overflow and cold-start bridge Public-class data only — policy-gated
Response path
Output guardrailsSchema, grounding vs. citations, leakage & PII redaction
Streaming responseProgressive tokens, cancellation propagation
Telemetry & meteringRouting, latency, tokens per tenant, outcome
Data, search, and event plane peer stores — no data flows between them directly; ingestion writes them, retrieval reads them
Object storageOriginals & versions · DR rebuild origin
Metadata DBTenants, live ACLs, versions, job state
Vector indexPer-tenant namespaces · blue/green aliases
Keyword indexExact terms · hybrid search partner
Queue / event busDomain events, jobs, retries, DLQs
Continuous RAG ingestion scales independently from user traffic; optimized for correctness over latency
SourcesUploads, S3, APIs, webhooks, CDC, polling
Ingestion queueBy reference, idempotent, DLQ-backed
Parse & classifyText, structure, sensitivity, content hash
Chunk & enrichSections, metadata, ACLs, versions
Embed & indexAtomic activation of the new version
ACL fast pathMetadata-only patch, skips re-embed Revocation SLO: seconds, not reindex time
Embedding migrationDual-index blue/green, full backfill, alias swap
Freshness & quality metricsLag, failures, deletions, recall, DLQ depth
Feedback & evaluation loop closes the loop from production behavior back into routing and prompts
Audit & tracingEnd-to-end traces, immutable audit log
Eval set curationSampled traces, user signals, regressions
Quality evaluationsGrounding, refusals, routing accuracy
Routing & prompt updatesThresholds, prompts, cache policy
↺ feeds back into the orchestrator's routing thresholds, prompts, and cache policies
Cross-cutting platform controls
Secrets & identity
Cost & capacity
Redaction & retention
DR & multi-region
Tenant isolation
Two planes with different consistency needs: the live path optimizes for latency and degrades gracefully; the ingestion path optimizes for correctness and versioned atomicity. The queue is the seam between them — and most operational risk lives where they touch (ACL staleness, freshness lag, index migration).