Phase 3 — Microsoft IQ (Foundry IQ retrieval)

This is the only phase that leaves the browser. Microsoft IQ binds a Copilot Studio agent to an Azure AI Search knowledge base; it does not create one. The knowledge base has to exist first, and building it requires Azure CLI and a REST surface that has no UI.

The goal is agentic retrieval — query planning plus answer synthesis — over docs.ray.io.

The add dialog (agent-iq-add-button) offers exactly one option: Foundry IQ (Preview).

Capture: 18_iq_add-dialog.png

Step 9a — Provision Azure AI Search

az account set --subscription 88dc9c38-361e-4516-9072-ae9504620c0f

az search service create `
  --name darbotlm `
  --resource-group rg-DarbotLM `
  --sku basic `
  --location centralus

az search admin-key show `
  --service-name darbotlm `
  --resource-group rg-DarbotLM `
  --query primaryKey -o tsv

Basic tier or higher is mandatory

A Free-tier search service's knowledge bases are never surfaced by the Copilot Studio picker. The picker reports "No knowledge bases found" with no explanation of why.

This was the key blocker of the entire build. A first service, raybot-foundryiq-search, was created on the Free tier in Australia East. Its knowledge base worked correctly when queried over REST and returned proper synthesised answers. It never once appeared in the picker. The service was discarded and rebuilt on Basic.

If the picker is empty, check the SKU before you check anything else.

Step 9b — Build the knowledge base

The /knowledgeBases endpoint exists only on api-version 2025-11-01-preview. Earlier preview versions return 404 for the whole path. There is no portal UI for this surface.

$svc = "https://darbotlm.search.windows.net"
$ver = "2025-11-01-preview"
$hdr = @{ "api-key" = $searchKey; "Content-Type" = "application/json" }

The knowledge source

A source of kind: "web" needs no further parameters.

PUT $svc/knowledgeSources/ray-docs-web?api-version=2025-11-01-preview

{
  "name": "ray-docs-web",
  "kind": "web"
}

The knowledge base

A web-backed knowledge base requires outputMode: "answerSynthesis" and an Azure OpenAI model for query planning. It will not accept extractive output over a web source.

PUT $svc/knowledgeBases/raybot-kb?api-version=2025-11-01-preview

{
  "name": "raybot-kb",
  "knowledgeSources": [{ "name": "ray-docs-web" }],
  "outputMode": "answerSynthesis",
  "retrievalReasoningEffort": { "kind": "low" },
  "answerInstructions": "...",
  "models": [
    {
      "kind": "azureOpenAI",
      "azureOpenAIParameters": {
        "resourceUri": "https://darbotlm-resource.openai.azure.com/",
        "deploymentId": "gpt-4.1",
        "modelName": "gpt-4.1",
        "apiKey": "$aoaiKey"
      }
    }
  ]
}

Verify before leaving Azure

POST $svc/knowledgeBases/raybot-kb/retrieve?api-version=2025-11-01-preview

A healthy response contains query-planning activity, web-search activity, answer-synthesis activity, and web references. Confirm this before returning to Copilot Studio — it isolates Azure-side problems from binding-side problems.

Step 9c — Attach in Copilot Studio

playwright-browser_click   agent-iq-add-button
playwright-browser_click   agent-iq-foundry-add-button

The connection form is backed by the azureaisearchfoundryiq connector and asks for two values: Search Service Endpoint and API Key.

Use the native value setter here

This form jumbles under type and fill exactly as the skill form does. Set both fields through browser_evaluate with the native value setter. See Phase 2.

Then Create, Next, and the "Select a knowledge base" step. That picker is populated by the Dataverse action AzureFoundryIQKnowledgeBases — which is the call that silently omits Free-tier services. Check raybot-kb, Add to agent, then Save and Publish.

Captures: 19_iq_foundry-connection.png, 20_iq_select-kb.png, 21_iq_attached.png

Step 9d — Verify in Preview

Ask a question that can only be answered from the documentation:

Explain Ray object spilling

The trace shows raybot-kb invoked, a "Searched knowledge" step, and a grounded answer with a Citations panel naming Object Spilling and Memory Management from Ray 2.55.1.

This is the proof that the binding works. It is also, as Phase 5 shows, the only surface that exercises this path — the Evaluate runtime does not.

Capture: 22_iq_verified-preview.png

Resources created

Resource	Detail
`darbotlm`	Azure AI Search, Basic, Central US
`ray-docs-web`	Knowledge source, kind `web`
`raybot-kb`	Knowledge base, answerSynthesis, reasoning effort low
`gpt-4.1`	Azure OpenAI deployment on `darbotlm-resource`
`raybot-foundryiq-search`	Free, Australia East. Discarded. Works over REST, invisible to the picker.

Captures

5 of 95 captures

18. Phase 3 - Microsoft IQAdd Microsoft IQ dialog (Foundry IQ)screenshots/18_iq_add-dialog.png

19. Phase 3 - Microsoft IQMicrosoft IQ - Foundry connectionscreenshots/19_iq_foundry-connection.png

20. Phase 3 - Microsoft IQFoundry IQ - select knowledge basescreenshots/20_iq_select-kb.png

21. Phase 3 - Microsoft IQMicrosoft IQ (raybot-kb) attachedscreenshots/21_iq_attached.png

22. Phase 3 - Microsoft IQIQ grounding verified in Previewscreenshots/22_iq_verified-preview.png

What Phase 3 establishes

Microsoft IQ binds rather than creates. The Azure side must be finished first.
SKU gates visibility, not function. A Free-tier knowledge base answers correctly and is still unusable, with no diagnostic in the UI.
API version gates existence. 2025-11-01-preview is the only version exposing /knowledgeBases.
Web sources force answer synthesis and require an Azure OpenAI deployment for query planning.
Verify at each boundary. retrieve over REST proves Azure; Preview proves the binding. Testing only at the end conflates two very different failure modes.

Step 9a — Provision Azure AI Search​

Step 9b — Build the knowledge base​

The knowledge source​

The knowledge base​

Verify before leaving Azure​

Step 9c — Attach in Copilot Studio​

Step 9d — Verify in Preview​

Resources created​

Captures​

What Phase 3 establishes​