RAG

Today: prompt-grounded answers + source rendering

The shipped path for grounding answers in your content is the system prompt: include your services, pricing, FAQ, and key page URLs there, and the bot answers from that knowledge. This works well for the small-to-medium knowledge bases the widget is designed for.

On the widget side, the source-display UI already exists: the chat API’s response type allows an optional sources array, and the widget renders a source icon with a badge count plus a slide-out sidebar (ChatSources) grouping links by type:

{
  "reply": "Our hourly rate is $75...",
  "sources": [
    { "url": "https://example.com/pricing", "title": "Pricing", "type": "page" }
  ]
}

type	Meaning
"blog"	Blog post
"page"	Site page
"external"	Third-party link

The bundled worker does not populate sources yet — if you run a custom backend that performs retrieval, you can return sources today and the widget will render them.

Planned: pluggable retrieval

• Pluggable retriever interface + Cloudflare Vectorize reference implementation — embed your content, retrieve relevant chunks per question, and ground replies in them — #52
• Inline citations and source cards rendered from RAG results — #56

Neither is shipped yet; the issues track design and progress.