PRDCT-353: fix the applications audit's live code-accuracy findings (Generative AI + trigger apps)

[Iamfle4ka]

The June audit flagged the swapped authentication as a blocker. It has been live ever since. Fixing it surfaced a worse defect on the same page, so this PR covers both.

1. The swapped key sources — the audit's finding

OpenAI Service Authentication told the reader to get their key from the Azure Portal; Azure OpenAI Service Authentication step 1 sent them to OpenAI platform settings. Each section now names its own provider.

The audit described this as a section swap, and acting on that would have been wrong. Only the key-source sentences were crossed — Azure steps 3–5 (API Base, Deployment ID, API Version) are genuinely Azure-only, each carrying options.dependencies.service = "azure_openai" in the live schema. Swapping the sections wholesale would have broken the half that was already right.

Confirmed two ways, independently of the audit:

• the vendor's own component_config/configuration_description.md in keboola/component-generative-ai;
• src/client/openai_client.py, where AzureOpenAIClient takes azure_endpoint/azure_deployment/api_version while the plain OpenAIClient takes only api_key.

The crossing entered in 0ff8c2aa ("replace OpenAI app docs with Generative AI app docs") — the page it replaced was correct. This restores that fact.

2. A reader failed at step one

The page said "set up a new configuration for the OpenAI application". But kds-team.app-open-ai is flagged deprecated and excludeFromNewList — it cannot be picked from the new-component list at all. The component this page actually documents is kds-team.app-generative-ai, named Generative AI, whose documentationUrl points here.

So anyone following the page hunted for a component they could not add. Fixed, with a note that existing OpenAI configurations keep working. That component also carries appInfo.beta, which the page never mentioned — now labelled.

3. Two providers documented, five shipped

authentication.service enum: openai, azure_openai, google, huggingface, anthropic. The lede now lists all five, including that Google is US-stack only and Hugging Face needs a per-row Endpoint URL.

4. UI labels, against the live configurationSchema

Page said	Actual field
API Base URL	API Base
List Models	List models
"Prompt Template Sync Action"	no such thing — it is the Prompt Templates section with a Prompt Template select

Also surfaced Maximum token spend (Optional), a prominent live option the page omitted.

5. Links

The Azure quickstart link now 301s to "Use the Responses API", which has no key-retrieval section — so step 1 pointed at guidance that had disappeared. Repointed to the Azure create-resource doc, which does cover key and endpoint. Every external link on the page was checked live.

Left for an owner rather than guessed

• VERIFY(owner) on the Deployment ID click-path: Microsoft now routes deployments through the Foundry portal (ai.azure.com → Deployments), not "Azure Portal → Azure OpenAI → Model Deployments". I removed the stale path rather than invent a new one — someone with a live Azure tenant should confirm before a click-path goes back in.
• platform.openai.com/account/api-keys returns 403 to any non-browser client, so its liveness cannot be verified from here. The page and the component's own config description use the same URL, so they at least agree.

Verification

Build clean, 306 pages. audit-phase2: 45 broken links / 0 missing images, and zero findings of any kind on this page.

Found by reconciling the June audits against the live sites; companion to #1080.

---

Second commit — the trigger apps

This closes the other two code-accuracy findings the June applications audit recorded on PRDCT-353. Both verified field-by-field against each component's live configurationSchema in the public Storage component index.

Power BI Refresh — the documented sample config could not have worked

Page said	Live schema
"datasets": [{ "dataset_input": "<uuid>" }]	dataset_list, and its items are plain UUID strings
"wait": true	string enum "Yes" / "No", default "No"
"alldatasets": false	string enum "Yes" / "No", default "No"
field "Fail if any dataset fails"	real label Wait for all datasets to finish
"Load Workspaces" / "Reload Dataset Names"	Load workspaces / Reload dataset names
"Interval" / "Timeout"	Refresh job status polling interval(s) / … timeout

Anyone who copied that JSON got a configuration the component does not read — wrong key, wrong item shape, and wrong types on two fields.

On the mislabeled option: the page's meaning was right (the schema's description is "End job with an error if any dataset finishes with a failed status") but no field by that name exists in the UI. Worth flagging to the component owner that alldatasets' own title and description disagree — per the 07-28 components call, that copy is fixed at the source, not here.

Also corrected the workspace hint to the schema's own wording, and added the defaults (30 / 3600) plus the project-level maximum timeout the schema mentions.

dbt Cloud Job Trigger — two parameters missing, and one made a sentence wrong

• ignore_artifacts ("Don't store artifacts", default false) was undocumented while the page stated unconditionally that the component "will also store all available artifacts". Now documented, and that sentence is qualified.
• dbt_cloud_url ("DBT Cloud URL", default https://cloud.getdbt.com) was undocumented — it is exactly what a non-default region or a single-tenant dbt Cloud instance needs.
• Smaller: the field is API Key; the text now names the Max wait time field rather than the raw max_wait_time parameter, and carries its default (60s) and the Cause default (Triggered from Keboola).

Left alone deliberately: the developers.keboola.com artifacts link stays until that section migrates into help, per the link rule — it is the one audit-phase2 finding on either page.

Verification

Build clean, 306 pages. audit-phase2 at 45 broken links / 0 missing images, unchanged from main.

Scope note: with this commit the PR covers all three live findings the audit recorded for components/applications/, which keeps it to one section = one PR rather than adding a third PR to a queue where unmerged work is already the bottleneck.

[linear-code]

PRDCT-353

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
connection-docs	Ready	Preview	Aug 5, 2026 4:26pm

[keboola-pr-reviewer-bot]

Verdict: needs_human (risk 3/5) · profile docs

Needs human: the diff makes many unverifiable product-behaviour claims across three component pages, and one external link's liveness is explicitly unverifiable.

Concerns:

• src/content/docs/components/applications/ai/generative-ai/index.md: Unverifiable product claims: OpenAI-app deprecation, five-provider enum, Google US-stack-only
• src/content/docs/components/applications/triggers/powerbi-refresh/index.md: Sample config rewritten to dataset_list/string enums — schema shape unverifiable from diff
• src/content/docs/components/applications/ai/generative-ai/index.md: platform.openai.com/account/api-keys liveness unverifiable (403 to non-browser clients)
• + 1 more concern — see Check Run summary.

Suggested reviewers: @keboola/docs

[keboola-pr-reviewer-bot]

Verdict: needs_human (risk 3/5) · profile docs

Escalating: the diff introduces many product-behaviour and config-schema claims across three component pages that can't be verified from the diff and need product-team confirmation.

Concerns:

• src/content/docs/components/applications/triggers/powerbi-refresh/index.md: Rewritten sample config (dataset_list, string "Yes"/"No" enums) unverifiable from diff; customers copy it.
• src/content/docs/components/applications/ai/generative-ai/index.md: New product claims: five providers, Google US-stack-only, OpenAI app deprecated, beta — need product confirmation.
• src/content/docs/components/applications/ai/generative-ai/index.md: VERIFY(owner) comment left in published content flagging an unconfirmed Azure click-path.

Suggested reviewers: @keboola/docs