Azure AI Search as a Knowledge Source
Semantic, vector, and hybrid search across document collections.
When your document collection outgrows simple search
Azure AI Search is like hiring a research librarian for your agent.
Copilot connectors and SharePoint knowledge work fine for thousands of documents. But when you have tens of thousands — or hundreds of thousands — of documents, and you need the agent to understand meaning, not just match keywords? That is when you bring in Azure AI Search.
It is a dedicated search engine that lives in Azure. You feed it your documents, it builds an index, and your Copilot Studio agent queries that index to find relevant answers. You get four search modes: keyword (exact match), vector (meaning-based), semantic (AI-ranked), and hybrid (best of all worlds).
The four search modes
Understanding these four modes is essential. The exam will present scenarios and expect you to pick the right one.
| Feature | How it works | Strengths | Weaknesses | Best for |
|---|---|---|---|---|
| Keyword search | Classic BM25 text matching — searches for exact terms and variations in the index | Fast, predictable, works well for technical terms, product codes, and exact phrases | Misses synonyms and conceptual matches — 'automobile' will not match 'car' | Structured queries, known terminology, product SKU lookups |
| Vector search | Converts query and documents to numerical embeddings — finds conceptually similar content | Understands meaning — 'vehicle maintenance' matches 'car repair schedule' | Requires an embedding model (Azure OpenAI), higher compute cost, embeddings must be generated at index time | Natural-language questions, cross-language search, finding conceptually related documents |
| Semantic ranking | AI re-ranker on top of keyword results — promotes the most relevant passages | Dramatically improves result quality for natural-language queries without needing embeddings | Only re-ranks existing keyword results — cannot find documents that keyword search missed | Improving keyword search quality when vector search is not feasible or not needed |
| Hybrid search | Runs keyword AND vector search in parallel, then fuses results with Reciprocal Rank Fusion (RRF) | Gets the best of both — exact matches from keyword, conceptual matches from vector | Most complex setup, highest compute cost, requires embedding model | Production workloads where query patterns are unpredictable — recommended default for most agents |
Exam tip: hybrid + semantic is the gold standard
For the exam, the recommended production pattern is hybrid search with semantic ranking. This gives you keyword precision, vector conceptual understanding, and AI-powered re-ranking in one query. If a question asks for the “best” or “recommended” approach for a Copilot Studio agent, hybrid + semantic is almost always the answer — unless constraints (cost, no embedding model) rule it out.
Index architecture — what you need to know
An Azure AI Search index is like a database table optimised for search. Understanding its structure helps you connect it correctly.
| Component | What it is | Why it matters |
|---|---|---|
| Index | A collection of documents with a defined schema | One index per knowledge domain (e.g., one for medical papers, one for HR policies) |
| Fields | Named attributes on each document (title, content, category, embedding) | Fields marked as “searchable” are included in full-text search. Fields marked “filterable” support pre-query narrowing |
| Indexer | An automated pipeline that pulls data from a source into the index | Blob Storage, Cosmos DB, Azure SQL, and other sources. Runs on a schedule or on-demand |
| Skillset | Optional AI enrichment pipeline attached to an indexer | OCR for scanned PDFs, language detection, entity extraction, chunking for vector search |
| Scoring profile | Custom relevance tuning rules | Boost recent documents, prioritise certain fields, weight by category |
Chunking matters for RAG
When using vector search, documents must be split into chunks (typically 500-2,000 tokens). Each chunk gets its own embedding vector. The integrated vectorisation feature in Azure AI Search can handle chunking and embedding automatically using a skillset with the Azure OpenAI embedding skill. The exam may reference “integrated vectorisation” — it means the search service handles chunking + embedding during indexing so you do not need a separate pipeline.
Connecting Azure AI Search to Copilot Studio
The connection workflow has specific steps the exam expects you to know:
- Create the Azure AI Search resource in your Azure subscription (choose the tier based on volume — Free, Basic, Standard S1/S2/S3).
- Create an index — define the schema, configure fields, optionally add a skillset for enrichment.
- Populate the index — run an indexer or push data via the REST API.
- In Copilot Studio, go to the agent’s knowledge sources and select “Azure AI Search.”
- Provide the connection details: search service endpoint URL, index name, and authentication.
- Configure field mappings — tell Copilot Studio which index fields contain the title, content, and URL for citations.
- Test — ask questions in the test pane and verify the agent returns grounded answers from the index.
Authentication options
| Method | How it works | When to use |
|---|---|---|
| API key | Pass the search service admin or query key in the request header | Simplest setup — good for development and testing |
| Managed identity | Copilot Studio environment uses a system-assigned managed identity with RBAC role on the search service | Production recommended — no keys to rotate, follows zero-trust principles |
| Microsoft Entra ID token | OAuth 2.0 bearer token from Entra ID | Advanced scenarios with fine-grained access control |
How this differs from Module 21 (Foundry RAG)
In Module 21, you will connect Azure AI Search through Azure AI Foundry to build a full RAG (Retrieval-Augmented Generation) pipeline with a Foundry model. In this module, the agent connects to Azure AI Search directly as a knowledge source — no Foundry model in between. The search results feed into Copilot Studio’s built-in generative answers capability. Same search service, different integration pattern. The exam distinguishes these two paths.
Scenario: Lena indexes 50,000 medical papers
Lena is the AI engineer at a healthcare analytics firm. Their clinical research team needs an agent that can answer questions across 50,000 published medical papers stored as PDFs in Azure Blob Storage.
Lena’s architecture: she creates an Azure AI Search resource (Standard S1 tier for the volume), builds an index with fields for title, authors, abstract, full_text, publication_date, and a vector field for embeddings. She configures an indexer with a skillset that includes OCR (some older papers are scanned images), text chunking (1,000-token chunks), and the Azure OpenAI embedding skill for vector search.
After the indexer runs, she connects the search index to the Copilot Studio agent using managed identity authentication. She maps the title, full_text, and a URL field for citations. In the test pane, she asks: “What are the latest findings on immunotherapy response rates in stage 3 melanoma?” The agent returns a grounded answer citing three specific papers with publication dates — powered by hybrid search (keyword for “melanoma” + vector for conceptual “immunotherapy response” matching) with semantic ranking.
50,000 papers, searchable in seconds.
Lena's healthcare agent needs to find papers about 'treatment efficacy for resistant tumours' even when papers use terms like 'therapeutic effectiveness' and 'refractory neoplasms.' Which search mode handles this best?
A developer connects Azure AI Search to their Copilot Studio agent but the agent never returns results from one specific field. What is the most likely cause?
Which authentication method is recommended for a production Copilot Studio agent connecting to Azure AI Search?
🎬 Video coming soon
Azure AI Search as a Knowledge Source