arcodash
Start free
PlatformConnectorsPricingSecurityDocsBlog
Sign inStart free

AI provider configuration

arcodash’s AI features (the “Ask with AI” SQL agent and Fix with AI, Auto Visualize, data source AI profile descriptions, Suggest checks, and alert explanations) are backed by an AI provider your organization configures: Gemini, OpenAI, or Anthropic. Everything is off until a key is configured, and configuring or changing it requires admin access.

[!NOTE] Most AI features send schema-level metadata. Sampled column values and alert investigation query results (the model is asked to aggregate them) are the exceptions, and both have a switch in AI Settings (see Data sent to AI). Data handling & privacy lists exactly what each feature sends.

Where the keys come from

You supply your own API key per provider, for Gemini, create one in Google AI Studio; for OpenAI and Anthropic, in each provider’s own console. Add it under AI Settings, this requires admin access. Each provider has its own key, and the AI features are unavailable until at least one is set. Generation is billed to your own provider account, not to arcodash.

Until a key is configured, the features stay hidden rather than showing buttons that can’t work. For each provider, AI Settings reports whether a key is in effect and where it came from (key_source): org for a key your organization set, env for a deployment-level key configured by the operator, or none.

Setting a key

Paste a provider’s API key into its card in AI Settings and save. It’s validated for a plausible shape client-side, then encrypted at rest with AES-256-GCM before it’s stored, the same encryption mechanism used for connector credentials. Nobody can read a configured key back out through the API; you can only replace it or clear it.

Clear the field and save to remove that provider’s key. With no keys left, the AI features turn off again until a new key is set.

Choosing a model

Each provider’s card has a Model selector: a curated list of models arcodash knows work well, plus a custom option for entering any model id the provider accepts. Leaving it alone uses the provider’s default. The model is set per provider, per organization.

For OpenAI only, the card also accepts a base URL override, useful for pointing at an OpenAI-compatible endpoint instead of OpenAI itself.

Provider priority and fallback

With more than one provider keyed, the Provider priority list controls which one is used: the first entry is the primary provider, and every AI request goes to it. Reorder the list to switch primary.

The providers below the primary matter only if Enable fallback is switched on: then a request that fails at the primary is retried against the next keyed provider in order. With fallback off, the rest of the list is ignored.

[!NOTE] When a fallback runs, that request is billed to the fallback provider’s own API key, worth knowing if your providers have very different pricing.

Testing a key

Use Test connection on a provider’s card to confirm a key actually works before relying on it, it validates the key without spending any generation quota. You can test a key you’re about to save, or re-test whatever’s currently configured. The result distinguishes:

  • Valid, the key works.
  • Invalid, the provider rejected it; it’s likely mistyped or revoked.
  • Quota exceeded, the key is valid but its quota is used up; generation will fail until quota resets.
  • Network error, arcodash couldn’t reach the provider at all; try again.

Data sent to AI

The Data sent to AI card in AI Settings has two admin-only switches:

  • Include sample values in AI prompts (on by default): when off, sampled column values and min/max are left out of every prompt. Column statistics are still used.
  • Explain alerts with AI (on by default): when a monitor fires, the model proposes up to three read-only queries (it is asked to aggregate, which is not guaranteed), arcodash runs them on the monitor’s data source, and the results are sent to your provider to write a short explanation. This requires the sample-values switch to be on and a data source with an AI profile. When off, alerts still carry a plain statistics line.

Rate limiting

Ask with AI, Fix with AI, Auto Visualize, and Suggest checks share one rate limit, counted per user, per minute. The platform default is 10 requests per minute; an organization can override it under Usage limit in AI Settings. Clearing the override falls back to the platform default. Profile descriptions and alert explanations run in the background and don’t count against it. Hitting the limit fails the AI call with a rate-limited error, it never blocks the rest of the product.

Where to go next