The AI layer that never goes down
Health-aware multi-provider routing, RAG grounding, prompt-injection protection, and tamper-evident audit logs — all in one orchestration layer.
Multi-provider failover
OpenAI, Anthropic, Gemini, and Ollama are all wired and health-monitored. When one degrades, traffic shifts automatically — within the same conversation turn.
Health-aware routing
Latency and error rates are monitored continuously. Degraded providers are deprioritized before they impact end users.
Automatic failover
Non-retryable errors or latency spikes trigger automatic rerouting to the next eligible provider within the same turn.
Per-plan model selection
Free → Claude Haiku 4.5. Starter → Sonnet 4.6. Professional adds Opus 4.7. Business expands Opus quota. Enterprise: custom models + BYOK.
What is your return policy for damaged items?
Items damaged on arrival can be returned within 30 days for a full refund. [Source: returns-policy.pdf §3.2]
RAG grounding with citations
Every AI answer is grounded in your uploaded content via semantic search. Citations are included so customers can verify — and so agents can audit.
Semantic search
Vector embeddings over your uploaded docs, FAQs, and product data. Similarity scoring with configurable threshold.
Source citations
Every grounded answer includes the source document and section. Customers can verify; agents can audit.
Fallback handling
When no grounded answer meets the threshold, the AI refuses to fabricate — it escalates or says it doesn't know.
Guardrails + prompt-injection protection
The AI orchestration layer filters and classifies every input before it reaches the model — protecting against injection attacks, jailbreak attempts, and off-topic requests.
Intent classification
Every inbound message is classified for intent before model routing. Off-topic or high-risk intents are gated with configurable handlers.
Injection detection
Prompt injection patterns are detected and blocked. The AI context is sanitized before each model call.
Content moderation
Harmful, abusive, or off-brand content is flagged before delivery. Human review can be triggered automatically.
Rate limiting
Per-tenant and per-IP rate limits prevent abuse. Quota overages are metered, not silently swallowed.
Tamper-evident audit log
Every AI decision, human action, and configuration change is logged with a cryptographic hash chain. Entries cannot be altered without detection.
| Timestamp | Event | Hash |
|---|---|---|
| 2026-06-02T14:32:01Z | conversation.escalated | a3f2b1c… |
| 2026-06-02T14:31:58Z | ai.response.grounded | d9e7c4a… |
| 2026-06-02T14:31:55Z | message.received | 8b3f2e1… |
Cryptographic audit log available on Business and Enterprise plans