Azure AI Engineer · AI-103
Domain 5 · 10–15%Implement information extraction solutions

Information extraction

Azure AI Search pipelines and queries, knowledge bases for agents, and document extraction. Closest to CCA-F RAG, with Azure's names for every stage.

PracticeD5 practice bank →

16 items, all formats, practice or timed exam mode.

ABuild the search index

5.1

The indexing pipeline

Data source → indexer → skillset → index: the indexer pulls content, skills enrich it, the index stores it.

From CCA-F: The retrieval half of CCA-F RAG, made concrete: chunking, embedding and storing are pipeline stages you configure, not code you write.

How it works

  1. Data source (Blob, SQL, Cosmos DB, OneLake…)
  2. Skillset: crack/layout → Text Split → embedding
  3. Index projections write one document per chunk

Must know

  • Skills: OCR, Document Layout, Content Understanding skill (GA), Text Split, Azure OpenAI Embedding, GenAI Prompt; custom logic via the Web API skill.
  • Indexers run on a schedule and pick up changes.
  • Index projections map one source file to many chunk documents.

Traps

  • The semantic ranker isn't a pipeline stage; it runs at query time.
  • Custom code in the pipeline = Web API custom skill.

Check yourself: d5-o1 · d5-n1 · Microsoft Learn

5.2

Fields and integrated vectorization

Field attributes decide what a field can do; vectors need a vector field, an embedding skill and a vectorizer.

From CCA-F: New for you. CCA-F treated the retriever as a black box; here you set the schema yourself.

How it works

  1. Mark fields searchable, filterable, sortable, facetable, retrievable; one key
  2. Add a vector field with dimensions and a vector profile (HNSW)
  3. Embedding skill at indexing + vectorizer at query time

Must know

  • Filters need filterable; facets need facetable.
  • HNSW = approximate and fast; exhaustive KNN = exact and slow.
  • Vector filters apply before the vector search by default (pre-filter).

Traps

  • Embedding skill without a vectorizer: text vector queries fail.
  • Changing a field's attributes usually means rebuilding the index.

Check yourself: d5-q27 · d5-y1 · Microsoft Learn

BQuery and connect to agents

5.3

Query types: keyword, vector, hybrid, semantic

Keyword finds exact words, vector finds meaning, hybrid does both, the semantic ranker reorders the top results.

From CCA-F: CCA-F said retrieval quality decides answer quality. This card is the menu of retrieval styles and when each wins.

How it works

  1. Exact IDs only → keyword (BM25)
  2. Paraphrase → vector
  3. Both → hybrid (RRF); best order → add query_type=semantic

Must know

  • Hybrid merges lists with Reciprocal Rank Fusion.
  • The semantic ranker reranks the top 50 and returns captions and answers.
  • VectorizableTextQuery sends text (needs a vectorizer); VectorizedQuery sends your vector.
  • Scoring profiles boost by field or freshness.

Traps

  • Semantic ranker can't find documents keyword search missed.
  • Pure vector search misses exact codes.

Check yourself: d5-q28 · d5-c1 · Microsoft Learn

5.4

Knowledge bases and Foundry IQ

A knowledge base retrieves across several sources for an agent, which connects to it over MCP.

From CCA-F: Like CCA-F's subagent that does research and returns findings: the knowledge base plans and runs retrieval, and the agent gets passages back.

How it works

  1. Create knowledge sources (index, Blob, OneLake, web…)
  2. Create a knowledge base over them
  3. Connect the agent to its MCP endpoint with the project's managed identity

Must know

  • Renamed from 'knowledge agents' (Nov 2025).
  • GA: extractive retrieval. Query planning, answer synthesis and reasoning effort: preview.
  • Tool: knowledge_base_retrieve.

Traps

  • Answer synthesis in a GA-only stem: wrong.
  • File search for organisational knowledge: wrong tool.

Check yourself: d5-q30 · d5-s1 · Microsoft Learn

5.5

Search access and security

Use Entra roles for search: one to query, one to load, one to manage.

From CCA-F: CCA-F's least-privilege tools principle, as Azure role names.

How it works

  1. Query → Search Index Data Reader
  2. Load documents → Search Index Data Contributor
  3. Create indexes, indexers → Search Service Contributor

Must know

  • Turn on role-based access on the search service; you can then disable keys.
  • Indexers reach data sources with the search service's managed identity.
  • Private endpoints and shared private links for private data.

Traps

  • Service Contributor manages objects but can't query data.
  • An agent's Search tool needs the project identity to have the right Search roles.

Check yourself: d5-n2 · Microsoft Learn

CExtract from documents

5.6

Choose: Content Understanding or Document Intelligence

Content Understanding for multimodal, RAG and schema-defined fields; Document Intelligence for standard forms, labelled custom models and offline containers.

From CCA-F: New for you. A service-choice question; the stem's constraint (no labels, offline, many layouts) picks the service.

How it works

  1. Offline or on-premises → Document Intelligence container
  2. Standard form (invoice, receipt, ID) → prebuilt model
  3. Varied layouts, no labelling, RAG → Content Understanding

Must know

  • Document Intelligence v4.0 is GA and not deprecated.
  • Custom template = fixed layout; custom neural = varied structured layouts (as few as 5 labelled samples).
  • Content Understanding uses your LLM deployments; read/layout don't need them.

Traps

  • "Newest service always" fails the offline case.
  • Custom neural still needs labelled samples.

Check yourself: d5-q29 · d5-y2 · Microsoft Learn

5.7

Content Understanding analyzers

Define the fields you want; the analyzer extracts, generates or classifies them, with confidence and source location.

From CCA-F: A strict output schema, as in CCA-F extraction, with grounding: each value points to where it came from.

How it works

  1. Base on a prebuilt (prebuilt-document) or start blank
  2. Add fields: name, type, method (extract / generate / classify)
  3. Analyze; read values, confidence and source

Must know

  • Classify fields take a fixed set of categories.
  • Confidence lets you route low-confidence results to a person.
  • Analyze is long-running: poll.

Traps

  • Generated fields aren't copied from the document: use extract when the value must appear verbatim.

Check yourself: d5-n3 · d5-c2 · Microsoft Learn