5.1 Azure AI Search — Fundamentals and Architecture

Quick Answer: Azure AI Search runs an indexing pipeline: Data Source -> Indexer -> Skillset -> Index -> Query. Indexers pull from Blob Storage, Azure SQL, Cosmos DB, ADLS Gen2, and SharePoint. Skillsets enrich data with Optical Character Recognition (OCR), Named Entity Recognition (NER), and key phrases. Vector search stores embeddings; semantic ranking re-ranks results. For RAG, use hybrid search (keyword + vector) with semantic ranking.

Where This Sits on AI-102

Knowledge mining and information extraction (Domain 6) is worth 15-20% of the AI-102 exam, which has roughly 40-60 questions, a 100-minute time limit, a passing score of 700/1000, and a $165 USD registration fee. Note for 2026: Microsoft has scheduled the Azure AI Engineer Associate (AI-102) exam to retire on June 30, 2026 — confirm the current edition before booking. Azure AI Search questions are heavy on case-study drag-and-drop: you order pipeline components, fix a broken field mapping, or choose the right search mode for a RAG scenario.

Indexing Pipeline Order

[Data Sources]            [AI Enrichment]            [Search]
Blob / SQL / Cosmos  -->  Skillset (OCR, NER,   -->  Index (full-text +
DB / ADLS Gen2 /          key phrases, custom)        vector + semantic)
SharePoint
      ^                                                     |
   Indexer  ----------------- coordinates ------------------+

The indexer is the orchestrator: it connects to one data source, optionally invokes one skillset, and writes documents to exactly one index. A common trap presents the components scrambled (Index first, or Skillset before Indexer) — the correct sequence is always Data Source -> Indexer -> Skillset -> Index.

Field Mappings vs Output Field Mappings

This is the single most-tested logistics detail in 5.1:

If a skill output never appears in the index, the missing piece is almost always an absent outputFieldMapping — not a skillset error.

Index Field Attributes

{"name": "id", "type": "Edm.String", "key": true, "filterable": true},
{"name": "content", "type": "Edm.String", "searchable": true, "analyzer": "en.microsoft"},
{"name": "category", "type": "Edm.String", "filterable": true, "facetable": true},
{"name": "contentVector", "type": "Collection(Edm.Single)", "searchable": true, "dimensions": 1536, "vectorSearchProfile": "vp"}

Trap: attributes are immutable once the index is created. Adding filterable to an existing field requires rebuilding the index — you cannot patch it in place. Only one field can be key, and it must be Edm.String.

Vector Search

Vector search retrieves by semantic similarity instead of exact tokens. You store an embedding (e.g., 1536 floats from Azure OpenAI text-embedding-ada-002, or 3072 from text-embedding-3-large) in a Collection(Edm.Single) field, then query for the k nearest neighbors.

"vectorSearch": {
  "algorithms": [{
    "name": "hnsw-1", "kind": "hnsw",
    "hnswParameters": {"metric": "cosine", "m": 4, "efConstruction": 400, "efSearch": 500}
  }],
  "profiles": [{"name": "vp", "algorithm": "hnsw-1"}]
}

The dimensions value must match the embedding model exactly — mismatching 1536 against a 3072-dim model throws an indexing error.

from azure.search.documents.models import VectorizedQuery
results = search_client.search(
    search_text=None,
    vector_queries=[VectorizedQuery(
        vector=query_embedding, k_nearest_neighbors=5, fields="contentVector")])

Hybrid Search and Semantic Ranking

Hybrid search runs BM25 keyword scoring and vector search in parallel, then fuses ranks with Reciprocal Rank Fusion (RRF). Layering the semantic ranker on top adds an L2 deep-learning re-rank of the top 50 candidates, plus semantic captions (the best passage per hit) and semantic answers (a verbatim extracted answer).

Semantic ranking requires a semantic configuration in the index and a Basic tier or higher (not free). For a RAG chatbot scenario, the exam-correct answer is almost always hybrid search with semantic ranking — it maximizes both recall and the relevance of the few chunks you feed the language model.

On the Exam: Distinguish BM25 (keyword), vector (embeddings, k-NN), and semantic ranking (re-rank). When a question says "customers misspell or paraphrase queries," the fix is vector or hybrid search; when it says "surface a direct answer snippet," enable semantic answers.