AI enrichment (experimental)

AI enrichment is an experimental capability that adds a human-readable summary to each change record. It is disabled by default, opt-in per cluster, and not part of the telark flagship surface. Treat it as a footnote, not a pillar.

If you do not need it, leave it off. The rest of telark is fully functional without it.

Where the toggle lives

Enrichment is gated by the GlobalConfig CRD:

spec:
  ai:
    enabled: false        # default
    provider: ""          # one of: groq | ollama | anthropic | gemini
    apiKey: ""            # required for the cloud providers

The dashboard surfaces this under Settings → Insights & governance → AI insights. Patching the field via the dashboard calls the exporter's GlobalConfig handler.

Supported providers

Four providers are wired today:

Provider	Notes
`groq`	Hosted. Requires `apiKey`. Optimised for latency.
`ollama`	Self-hosted. Needs an Ollama endpoint reachable from the discovery deployment. No API key.
`anthropic`	Hosted. Requires `apiKey`.
`gemini`	Hosted. Requires `apiKey`.

The provider is selected by the provider field. Switching providers is a hot operation — the next change cycle picks up the new selection.

What gets sent

telark sends the structured change record (class, severity, generation, diff summary, affected resource refs) plus a short prompt. It does not send manifest payloads, secret contents, or anything outside the change record. The exact prompt template lives in the discovery service's enrichment helpers.

If the data sent is more than your compliance posture allows, the right answer is to leave the feature disabled. The telark flagship features (discovery, protection plans, snapshots, rollback) do not depend on enrichment.

What you get back

The enrichment output is a short paragraph that goes onto the change cycle record as an enrichment field. The dashboard renders it inline on the change-history panel of an application detail page when present. If enrichment fails or is disabled, the record is rendered without it — there is no fallback noise.

Failure handling

Enrichment is best-effort. If the provider call fails, times out, or returns a malformed response, the change record is persisted without an enrichment field and the cycle continues. The failure is logged but does not block discovery, protection plans, or any other surface.

What this is not

Not on the landing page. The marketing surface does not list AI enrichment as a pillar. It exists for teams that want it, not as a selling point.
Not multi-tenant. A single provider + key configuration is shared across the whole telark install.
Not an audit feature. Outputs are descriptive, not authoritative. Do not feed them into automation that takes action on the cluster.
Not a guarantee. Hosted providers can rate-limit, change behaviour, or be deprecated. Self-host with Ollama if predictable behaviour matters.