Rich content trust

Rich content — product cards, offer carousels, media tiles — renders in the widget only when it originates from a successful, verified tool execution. The LLM cannot fabricate structured cards from plain text.

Trust model

Source	Renders in chat?
Tool result with valid `result_transform`	Yes
Tool result without transform (raw JSON)	No — text summary only
LLM-generated JSON in message text	No — stripped
Knowledge base citations	Text + citation links only
Failed tool call	Error message; no cards

This prevents prompt-injection attacks that try to inject fake product listings or phishing links as structured UI.

Pipeline

flowchart LR
  A[User message] --> B[LLM selects tool]
  B --> C[Tool executes against your API]
  C --> D{Success?}
  D -->|Yes| E[result_transform maps JSON to card schema]
  D -->|No| F[Text error — no rich content]
  E --> G[Sanitize URLs and fields]
  G --> H[Render in widget]

result_transform

Tools declare how API JSON maps to card fields:

{
  "result_transform": {
    "type": "carousel",
    "mapping": {
      "items": "$.offers",
      "title": "$.name",
      "image_url": "$.image",
      "link_url": "$.booking_url"
    }
  }
}

Only fields defined in the transform schema are rendered. Unknown fields are dropped.

URL sanitization

Check	Action
`javascript:` URLs	Blocked
Non-HTTPS links (production)	Blocked unless allowlist configured
Off-domain images	Allowed if HTTPS; flagged in logs
HTML in text fields	Escaped — no inline HTML in cards

Configure an optional image/link domain allowlist under Tools → Rich content → Trust policy.

Visitor visibility

Visitors never see:

Raw tool API responses
Credential headers or tokens
Internal tool error stack traces

They see user-friendly text and sanitized cards.

Dashboard verification

Use Tools → Test to preview rich content output before attaching to an assistant. The preview shows exactly what the widget will render.