Model provider controls
Some reviews ask for evidence that a particular provider's models can be turned off. Three layers exist, listed from the one you can see in the product to the strongest.
1 · Per-feature model selection in your workspace
Workspace settings → Models exposes one model selector for each of four features: Theme Detection, Audiences, Alerts and Narrative Testing. Each lists the models in the catalogue, from Google, OpenAI and Anthropic, plus any model endpoint the workspace has added. Three selectors default to a Google model and Narrative Testing to an Anthropic model.
A selector sets the model for its feature's main calls. Supporting calls a feature makes alongside them (signal extraction, document tagging, audience chat and model-side web search) use the platform default model, and move only when a selector names a model endpoint the workspace added itself. One setting brings Gemini back for an endpoint: with Workspace settings → Policy → Web search for your own model endpoints on (it is off by default), each web search the endpoint's model asks for runs as a Gemini call that receives the search query and nothing else, on Kumkuat's Gemini account or on the workspace's own Gemini key when it has added one. Choosing models therefore does not by itself remove a provider from a workspace; the per-call audit in section 3 shows which providers a workspace's calls actually reached.
2 · Platform-level credential removal
Provider credentials are held in Google Secret Manager and injected into the running service. Removing a provider's credential binding means no call to that provider from any workspace can authenticate, regardless of any model setting. Kumkuat can apply this fleet-wide if you require a hard control rather than a configuration one.
3 · Audit trail per call
Every model call is recorded with the workspace, the user, the feature, the provider and the model. Kumkuat can produce, on request, a per-provider count of calls for your workspace over any period. That record is the evidence that the configuration held, not just that it was set.
provider calls
gemini …
openai …
anthropic 0
Evidence Kumkuat can supply
| Evidence | Contents |
|---|---|
| Model selector screenshot | Workspace settings → Models, showing the four selectors and the models chosen |
| Per-provider call audit | The per-provider call count for your workspace over the evaluation period, drawn from Kumkuat's usage log |
| Written confirmation | A statement of the model each selector names and of the calls that use the platform default, with the audit on request |
Related controls
- Web grounding can be switched off per workspace so persona chat answers only from stored context; the other places a feature may search, and the switch for each, are on Web grounding.
- Connected libraries can be excluded from tagging and embedding so proprietary text is never sent to a provider unless a user submits it. See Knowledge bases.