Control what the model receives
Use this guide before connecting business data. Disclosed fields are the columns your application permits the assistant to describe or use. Excluding a field narrows assistant access; your server still owns authorization.
For a first integration, keep row samples, conversation memory and result follow-ups disabled. Enable each only after deciding what information your backend and model should receive.
What leaves the browser
Section titled “What leaves the browser”An interpretation request contains the prompt, disclosed field metadata, current disclosed filter/sort criteria, locale, timezone and reference instant. With grid actions enabled it also includes projected column layout and grouping. A standalone gridState callback can additionally disclose selected range coordinates. It does not send full column definitions, renderers, execution callbacks or row values by default. Prompts and filter operands may themselves contain business data.
Optional disclosures are independent:
samplesrequires an explicitgetRows(signal)callback and field allowlist. Values are projected to JSON scalars. Defaults are five rows and 4 KiB; hard limits are 50 rows and 16 KiB. Redact in your callback. The reference backend ignores browser samples.conversationretains complete prior prompt/summary pairs: default eight messages and 8 KiB, hard limits 20 messages and 16 KiB. Raw action data/details and previews are excluded. Visible answer text may still contain business data. Reset, effective field-policy changes and disposal clear this memory; failed, cancelled and stale turns do not enter it.answerFromAction.selectexplicitly chooses which completed immediate-action result can return to the model. Nothing is selected automatically.
Requests larger than 64 KiB are rejected before interpretation. The server must enforce its own field, action and data policy; a browser-provided descriptor is not authorization. AI exclusions also do not configure redaction in your exporter, chart or business service.
Next: Answer questions about data.