Practice Exams:

Microsoft AI-103: Operating Azure AI Search for RAG

AI & Machine Learning

Adding vectors to a search index does not automatically produce trustworthy retrieval-augmented answers. The application still has to ingest documents, attach meaningful metadata, apply access controls, choose retrieval methods, and prove that relevant evidence reaches the model. Azure AI Search combines lexical and vector retrieval, filters, semantic ranking, and indexing pipelines, making it a practical retrieval layer for Microsoft Foundry applications. Those capabilities work only as well as their configuration and underlying sources. A production design should explain which documents can be returned, how current they are, why the query selected them, and what the application does when no sufficient evidence exists.

On this page
  1. Treat the index as a governed data product
  2. Decide how content will be split and embedded
  3. Combine exact terminology with semantic similarity
  4. Apply semantic ranking where it adds value
  5. Enforce permissions before evidence reaches the model
  6. Choose ingestion and update strategies deliberately
  7. Compare classic and agentic retrieval approaches
  8. Evaluate retrieval separately from generation
  9. Keep the retrieval layer observable and recoverable

Treat the index as a governed data product

Begin with a source inventory, not an index wizard. Identify document owners, update frequency, access rules, version identifiers, and the operational consequences of stale content. A human-resources handbook, engineering runbook, and marketing FAQ should not automatically share one freshness policy or retrieval authorization scheme. Specify how each source enters the index, what constitutes an authoritative version, and how a deleted document is removed from search. The index is a derived system of record; it must not become the only surviving copy of the organization’s policies or files. A Microsoft AI-103 retrieval system should test whether Azure AI Search returns the right authorized evidence before measuring the fluency of the final answer.

Include source IDs, section titles, effective dates, content type, tenant boundaries, and permission attributes in the data model where these matter. A retrieval pipeline cannot recover reliable metadata that was discarded during ingestion. Choose fields that support filtering and diagnostics, not simply a large JSON blob for every document. Keep source-reference information sufficient for readers to open the original material they are authorized to view. The model should receive a passage because it is relevant and permitted, not merely because the vector similarity score happened to be high.

Decide how content will be split and embedded

Chunking establishes the units a search system can retrieve. Splitting every document into arbitrary character windows can detach a table from its heading or a policy exception from the clause it qualifies. Respect sections, document structure, and page boundaries where useful; overlap adjacent chunks only enough to preserve continuity. Test what happens when a real answer spans two chunks. A chunk may need concise contextual metadata describing its document and section so it remains understandable after being retrieved on its own.

Embeddings translate text into vectors for similarity search. Select an embedding model according to supported languages, domain terminology, dimensionality, indexing constraints, and observed retrieval results. Every stored vector must be generated with compatible model settings; silently mixing incompatible embedding generations in one field can damage retrieval. Record embedding model and index schema versions, and plan how a future embedding upgrade will rebuild or migrate the index. The right dimension is the one that meets quality and operational objectives, not automatically the largest offered value.

Combine exact terminology with semantic similarity

Lexical search remains essential when queries contain error codes, product SKUs, contract identifiers, or other exact strings. Vector search can help when a user describes the same concept using different words. Azure AI Search hybrid queries combine these methods, and Microsoft documents Reciprocal Rank Fusion as a method of merging candidate rankings. This is useful for enterprise questions that mix specialized terms and natural descriptions. It also creates tuning decisions: a hybrid setup needs enough lexical and vector candidates to expose the correct passage before later reranking.

Consider a service-desk question about a specific authentication error. A vector-only query might find broad identity guidance while missing the precise error message. A keyword-only query may find the message but miss the updated troubleshooting section that uses different terminology. A hybrid candidate set can surface both. Test across representative question types rather than declaring hybrid retrieval universally superior. Record which retrieval component found the relevant document and whether unnecessary candidates consumed context or confused the generator.

Apply semantic ranking where it adds value

Semantic ranker evaluates and reorders an initial set of retrieved text candidates. It cannot recover a relevant document that never made the candidate set, and it should not be mistaken for an authorization filter. Ensure the index exposes useful text fields and that the preceding retrieval stage supplies suitable candidates. Measure whether reranking improves evidence placement for the actual questions users ask. If a team only measures final generated answers, it may miss a ranking regression that the language model happens to conceal for some queries.

Reranking introduces latency and additional service behavior, so test it against measurable benefit. An application answering simple exact-ID lookups may need less sophisticated ranking than a legal-search assistant interpreting long requests. Keep the query pipeline visible in traces: keyword matches, vector matches, fusion results, final ranked passages, and cited sources. That visibility allows engineers to determine why an outdated procedure outranked a newer one, or why a semantically related document displaced the only passage that contained the answer.

Enforce permissions before evidence reaches the model

Search security cannot be left to a prompt that tells the model to ignore confidential documents. Enforce the actual access boundary through supported authorization features, security filters, or application-layer checks validated against the authenticated user. Record which identity the search service uses to access underlying data and what scope that identity has. A shared service principal that can retrieve every tenant’s documents requires carefully designed user-level authorization before any result becomes model context. The language model cannot reliably undo a leakage after text has already been supplied to it.

Permissions also change over time. If a document becomes restricted, update or invalidate corresponding index access metadata and any result caches. Include tests where two users issue identical queries but are entitled to different sources. A cached retrieved passage must not cross those boundaries simply because its content is relevant. Check whether document-level permission behavior is supported for the particular storage and search configuration in use; Microsoft features and integration limits vary. The desired result is that relevance is evaluated only among evidence the user is permitted to receive.

Choose ingestion and update strategies deliberately

Azure AI Search provides ways to index and enrich content, including integrated vectorization for suitable pipelines. Such features can simplify the data flow, but a production system still needs ownership of source updates, partial failures, schema migrations, and deletion behavior. Decide whether the source is pushed by an application, pulled through supported indexers, or processed by a separate ingestion workflow. Record the point at which a document becomes visible to queries so the user-facing application can explain freshness instead of assuming that upload means instant availability.

For mutable sources, use stable document identities and version-aware updates. A failed ingestion run should not leave a mix of new metadata and old vectors without alerting operators. Detect source removals, retry bounded failures, and reconcile counts between the source system and the index. For a regulated document, keep the original artifact and extraction provenance even when searchable text is updated. The index is an operational derivative, so its recovery plan should include deterministic rebuilding from versioned sources rather than relying only on a live backup of its current state.

For sources that are not already clean text, an Azure AI Search indexer can apply an enrichment skillset before indexing. Built-in skills can perform OCR, image analysis or language detection, while a custom Web API skill can run an approved domain-specific transformation. Consider an equipment manual that mixes paragraphs with labeled diagrams: an indexer can extract image text, merge it with surrounding content, preserve the source document identifier and map the resulting fields into the search index. The indexer must have valid input/output field mappings and the necessary Foundry resource configuration; image extraction and enrichment can add cost. Verify ingestion errors, freshness, access filtering and the resulting evidence in sample queries, rather than assuming an enriched field is automatically accurate or safe to share.

Compare classic and agentic retrieval approaches

A classic RAG application constructs search queries and chooses passages in application code before invoking a model. This provides predictable control over query construction and source filtering. Agentic retrieval patterns may allow a planning component to decompose or refine questions, which can be valuable for multi-part research. That flexibility increases the need to inspect tool choices, query counts, and permission handling. A complex retrieval planner should not be adopted merely because it sounds more advanced; the gain must be evident in answer quality for tasks that require it.

An employee policy assistant may need only a scoped hybrid query and a short grounded answer. A research workflow comparing several technical specifications may benefit from multiple targeted queries and an evidence synthesis step. Test both designs against the same information needs, then compare task completion, coverage, latency, token use, and error cases. Whatever retrieval strategy is chosen, every returned source still needs a stable reference and the same access controls. More autonomous planning does not change the requirement that evidence be current and authorized.

Evaluate retrieval separately from generation

Create a dataset of representative questions with known relevant passages and expected permission constraints. Track candidate recall, ranking position, freshness, and the presence of unauthorized results. Include queries that should return no evidence and cases where two documents conflict. Evaluate whether the search stage provides enough information for a correct answer before measuring the model’s prose. A hallucinated answer can originate in weak retrieval, but a model can also ignore strong retrieved evidence; measuring the two components separately identifies the responsible layer.

Use production traces to discover new retrieval failures. If users repeatedly ask about a service update that is absent from the index, improving embeddings will not fix ingestion. If exact device codes fail to match, investigate lexical configuration. If relevant candidates are present but poorly ordered, review ranking. Document every change to chunking, embeddings, filters, and search configuration as part of the release. This turns retrieval quality into an engineering outcome that can be maintained rather than a mysterious property of the model.

Keep the retrieval layer observable and recoverable

Monitor indexing delays, query error rates, unauthorized-result checks, average and tail latency, and search-service consumption. A rise in model token usage may reflect overly broad retrieval rather than a model change. Correlate query traces with the answer and its citations, but avoid logging full private documents without an explicit data policy. Keep the pipeline’s configuration reproducible so a schema update or deployment failure can be recovered from known sources and versioned settings.

A successful Azure AI Search design does not stop at a relevant top result in one demonstration. It provides a defensible source-to-answer path that honors permissions, tolerates updates, and explains failures. Hybrid and semantic retrieval, embedding fields, filters, and indexing tools are ways to build that path. Their value is measured by accurate evidence reaching the application consistently, with enough transparency for operators to diagnose when it does not.

Continue learning

Related guides

Vector, Hybrid, and Semantic Search for Better GroundingChoose keyword, vector, hybrid, and reranking strategies for RAG.Choosing Embeddings on AzureEvaluate embedding models for domain relevance, latency, and reindexing.RAG on Azure: Retrieval Quality Is the ProductMeasure retrieval precision, source eligibility, and evidence grounding.Hybrid Search in Azure AI SearchCombine lexical and vector retrieval with effective query ranking.

Related Posts

• Mastering the Basics: Your Guide to Microsoft Azure AI

• The Ultimate Overview of Microsoft AI-900: What You Should Know

• How to Use Microsoft Copilot: Your Guide to Boosting Productivity with AI

• Microsoft AI-300: From Experiment Tracking to an Auditable ML Lifecycle

• Microsoft AI-300: How to Roll Back a Model Without Rolling Back the App

• Microsoft AI-103: Safe Tool-Using Agents

• Microsoft AI-103: Hybrid Search and Reranking

• Microsoft AB-900: The Microsoft 365 Admin Role Now Includes AI

• Microsoft AI-901: Computer Vision, Language, and Generative AI

• Microsoft AI-103: Identity Boundaries for Foundry Agents