2.1.3. Azure AI Search
💡 First Principle: Azure AI Search is the bring-your-own-index option: you own the retrieval pipeline — chunking, vectorization, ranking — and Copilot Studio simply sends its queries to your index. Maximum control, maximum responsibility.
The prerequisites are where exams live. The index must be a vector index built with integrated vectorization — the "Import and vectorize data" path in the Azure portal — because at runtime, the same embedding model that vectorized your documents vectorizes the incoming question, and similarity search does the rest. Semantic ranking is supported but must be configured on the Azure side before you connect. Citations require the index to expose a URL field (or the metadata_storage_path field) so answers can link back to sources. And each connection points at exactly one vector index.
Connections authenticate with an admin key, service principal, client certificate, or Microsoft Entra ID — and the connection must be created through the knowledge UI, because manually improvised connections can corrupt the Azure AI Search dialog for every agent in the environment. Private endpoints and VNet-isolated search services are supported, which is frequently the reason this option gets chosen: the corpus that isn't allowed anywhere near a shared index.
⚠️ Common Misconception: Using Azure AI Search means building your own RAG pipeline in code. You build the index; Copilot Studio handles retrieval against it natively. The division of labor is exactly: your index, their orchestration.
⚠️ Exam Trap: If citations aren't appearing for Azure AI Search answers, check the index schema — a missing URL/metadata_storage_path field is the classic cause. The fix lives in Azure, not in Copilot Studio.
Reflection Question: Your legal team's corpus lives in a VNet-isolated Azure AI Search service, already vectorized. Which requirements in this scenario eliminate Copilot connectors, and which eliminate Power Platform connectors?
