Answers
Frequently asked questions
Straight answers about how the governed AI pipeline works. Anything not covered here is worth asking the discovery console.
AI Governance
Does AI-generated content get published automatically?
No. Every AI output is labelled as an AI GENERATED DRAFT and enters the same review queue as human drafts. Publication requires an explicit approve action from an editor, recorded in the audit log with their identity and the exact revision they approved.
Who can approve content for publication?
Approvers come from an allow-list configured through environment variables. Approval requires authentication, and the workflow enforces that the approver cannot be the same identity that generated the AI draft in the same session for regulated categories.
AI Discovery
Where do the answers in RAG search come from?
Answers are generated only from passages retrieved by the platform's own search index. Each response displays its sources with links to the underlying content. When retrieval finds nothing relevant enough, the assistant says so instead of answering from general knowledge.
About This Platform
Is this platform a real production system?
It is a technical demonstration built to production engineering standards. The architecture, workflows, APIs and tests are real implementations; demo-mode telemetry and some AI responses are clearly-labelled simulations so the experience works without any API keys.
AI Engineering
Which AI providers does the platform support?
The provider layer speaks the OpenAI-compatible chat completions protocol, so DeepSeek, OpenAI, Azure OpenAI or any compatible endpoint work by configuration alone - no code changes. A built-in simulated provider covers demo mode when no key is configured.
Security
How is prompt injection handled?
Retrieved content is treated as untrusted data, never as instructions. System prompts are separated from context blocks, retrieved text is delimited and neutralised, and outputs are validated against expected schemas before display. Known attack strings are additionally screened at ingestion.
Architecture
Can the platform work with Contentful?
Yes. The CMS layer is defined as an interface with two adapters: a local file-backed adapter used in demo mode and a Contentful Delivery API adapter for production. Swapping them is environment configuration; no application code changes.
Reliability
What happens when the AI provider is down or over budget?
Circuit breakers fail closed: the assistant reports temporary unavailability rather than degrading into ungrounded behaviour, cached results keep read paths working, and the observability dashboard shows the breaker state and recent error rates.