Under the hood

Built properly. Here are the details.

For the technical folks: what runs where, and how to plug in.

How a message flows

Channels

Website widgets · WhatsApp (soon)

Nubo engine

model routing · retrieval · reranking · tools · approvals

AI models

GPT · Claude

Your systems

MCP · REST actions

<!-- Paste into your site's HTML -->
<script
  src="https://qnragstorage.blob.core.windows.net/widget/widget.js?widgetId=YOUR_WIDGET_ID&pos=r">
</script>

<!-- That's it. The widget only loads on the
     addresses you allow for it. -->

One line on your site. Copy it from the widget's Install tab.

Models: OpenAI GPT and Anthropic Claude, with per-question auto-routing
Retrieval: Chunking strategies, vector search, optional cross-encoder reranking
Tools: MCP servers and HTTP actions: JSON Schema inputs, response allowlists, approval gates
Streaming: Replies stream token by token over server-sent events
Security: Signed widget tokens, per-widget allowed addresses, CSP frame-ancestors, secrets in Key Vault
Public API: REST with API keys, 60 requests a minute per key
Evaluation: LLM-as-a-judge, four metrics, scheduled runs, email alerts