9.3 Connect to Azure AI Search

9.3 Connect to Azure AI Search

Quick Answer: Prepare an Azure AI Search service with a vectorized index (integrated vectorization recommended), optional semantic ranker, then in Copilot Studio Add knowledge → Azure AI Search using a formal data connection. Prefer Entra ID–based auth patterns, map one vector index, validate citations via URL fields, and use Search when hybrid retrieval quality, scale, or network isolation exceed simple SharePoint file knowledge.

Azure AI Search is the third enterprise knowledge pillar on the AB-620 outline. Later Domain 2 also covers generative answers using Azure AI Search with Foundry—this section focuses on connecting Search as knowledge for agents in Copilot Studio.

Why Azure AI Search exists in the knowledge toolbox

SharePoint knowledge, uploaded documents, Copilot connectors, and Power Platform real-time knowledge each solve parts of enterprise grounding. Azure AI Search is the purpose-built search engine for large document collections with AI retrieval features:

• Full-text (keyword) search
• Vector search over embeddings
• Hybrid search combining both in one request
• Optional semantic ranker for relevance
• Controlled indexing pipelines, chunking, and embedding models
• Enterprise networking (including private endpoints)

Microsoft Learn positions Copilot Studio support around vectorized indexes using integrated vectorization: prepare data, choose an embedding model, use Import and vectorize data, and the same model can vectorize the incoming prompt at runtime—reducing custom glue code.

Indexes, hybrid search, and ranking

Exam trap: “Hybrid” is not marketing fluff—it is a concrete retrieval mode. Stems that complain about poor relevance on synonym-rich or multi-language content often want vector or hybrid + semantic ranker, not “add another SharePoint URL.”

Connecting Azure AI Search in Copilot Studio

Microsoft’s how-to emphasizes a formal data connection:

1. Open the agent.
2. Add knowledge from Overview, Knowledge, or generative answers properties.
3. Featured → Azure AI Search.
4. Create new connection.
5. Select Authentication type: Access Key, Client Certificate Auth, Service principal (Microsoft Entra ID application), or Microsoft Entra ID Integrated.
6. Enter required details (for key-based paths: Search endpoint URL and admin key).
7. Create connection (green check).
8. Next, enter the vector index name—only one vector index in this add path.
9. Add to agent. Status In progress while metadata indexes, then Ready.
10. Test knowledge and review citations.

Critical operational warning from Microsoft: Do not manually configure unsupported endpoint/API key wiring outside the supported data connection experience. A faulty Azure AI Search data connection can prevent the connection dialog from loading and is managed at environment level—affecting agents until you recover by resetting external access or recreating the agent, then re-adding with Data sources → Azure AI Search using proper Entra ID authentication patterns rather than broken manual setups.

Citations and grounding quality

Grounding quality is not only “did the model speak.” For Azure AI Search knowledge:

• Include a URL field with the real document link in the index so Copilot Studio can return citations.
• When metadata_storage_path exists, Copilot Studio can treat it as the citation; otherwise a field containing a complete URL may be used.
• Users must have permission to open cited destinations; restricted URLs produce frustrating dead ends even when retrieval succeeded.
• Pair retrieval quality (chunking, hybrid, semantic ranker, clean metadata) with agent settings that discourage ungrounded inventing when enterprise accuracy matters.
• Clear knowledge source descriptions help generative orchestration select Search among many sources.

Security trimming, identity, and networking

“Security trimming” means users only retrieve content they are allowed to see. Patterns vary by architecture:

Exam trap: Pointing citations at locked blob storage without granting users access “trims” the experience into failure. Retrieval security and citation accessibility both matter.

When to use Azure AI Search vs SharePoint files (and friends)

Azure AI Search is not “always better SharePoint.” It is the right answer when retrieval engineering, scale, hybrid quality, or Azure networking dominate. SharePoint knowledge remains correct for straightforward Microsoft 365 document grounding with user Entra auth.

Architecture scenario

Fabrikam stores product manuals, safety PDFs, and historical change notices across Azure Blob and a legacy CMS. Simple SharePoint attach cannot represent the corpus. Search engineers create an Azure AI Search service, run integrated vectorization, enable semantic ranker, and store metadata_storage_path for each chunk. Power Platform admins enable VNet support; Search uses a private endpoint. The agent builder adds Azure AI Search knowledge with Microsoft Entra ID Integrated authentication, selects the vector index, and tests questions such as “What PPE is required when servicing model XR-200?” Citations open the authorized PDF. High-risk “create maintenance work order” remains a tool, not a Search write.

If relevance is weak, the team does not first swap models randomly in the agent—they inspect chunk sizes, hybrid configuration, semantic ranker enablement, and whether the wrong index fields are searchable.

Relationship to Foundry (preview of later skills)

AB-620 also measures generative answers using Azure AI Search with Foundry. Conceptually: Search remains the retrieval substrate; Foundry contributes model and agent patterns. For this leaf, master connecting Search as knowledge—indexes, hybrid quality, auth, citations, and when Search beats file-only sources. Do not invent a requirement that every Search knowledge agent must also be a full Foundry multi-agent solution.

Pitfalls

1. Manual broken data connections at environment scope.
2. Keyword-only indexes when hybrid was required for meaning-based questions.
3. Missing citation URL fields.
4. Private Search service without Power Platform VNet alignment.
5. Using Search knowledge for transactions.
6. Choosing Search for a ten-file SharePoint FAQ (unnecessary complexity).
7. Ignoring user permissions on citation targets.

Success checklist

• Can describe vector index + hybrid + semantic ranker.
• Can walk Add knowledge with formal data connection and one vector index.
• Can argue Search vs SharePoint vs Copilot connectors vs PP real-time.
• Can explain citations, auth options, and private endpoint considerations.
• Treats Search as knowledge grounding, not a write API.

With Copilot connectors, Power Platform connectors, and Azure AI Search all fluent, you can design enterprise knowledge for AB-620 without falling for the universal exam trap: knowledge grounds; tools act.