Implement information extraction solutions

5.1 Build retrieval and grounding pipelines

The four Azure AI Search objects

Every indexing question comes down to which of these four you change:

flowchart LR
  DS["Data source<br/>Blob · ADLS Gen2 · Cosmos DB<br/>Azure SQL · Table · OneLake"]
  DS --> C1["1 · Document cracking<br/>open the file, extract<br/>text and images"]
  C1 --> C2["2 · Field mappings<br/>source field → index field"]
  C2 --> C3["3 · Skillset execution<br/>OCR · layout · chunk · embed<br/>enrich · custom skills"]
  C3 --> C4["4 · Output field mappings<br/>enriched node → index field"]
  C4 --> C5["5 · Push to index"]
  C3 -.-> KS["Knowledge store<br/>(non-search scenarios)"]
  C3 -.-> CACHE["Enrichment cache<br/>(reuse expensive OCR)"]

Enrichment: built-in and custom skills

AI enrichment is image and natural-language processing inside the indexer pipeline, extracting text and structure from content that could not otherwise be indexed. The skills the exam is most likely to name:

Chunking and integrated vectorization

flowchart TD
  D["Source document"] --> S{"Structure?"}
  S -- "Long prose" --> T["Text Split skill<br/>sentences or pages,<br/>with overlap"]
  S -- "Headings, tables,<br/>sections matter" --> L["Document Layout skill<br/>or Azure Content Understanding skill<br/>layout-aware chunking"]
  S -- "Scanned / image-only" --> O["OCR first,<br/>then chunk"]
  T --> E["Embedding skill"]
  L --> E
  O --> E
  E --> I["Index: chunk text +<br/>vector + parent key + metadata"]

Integrated vectorization extends both pipelines so you never compute embeddings yourself:

• During indexing — a chunking skill plus an embedding skill in the skillset, with vectors routed to dedicated vector fields.
• At query time — a vectorizer defined in the index and attached to vector fields through a vector profile, so you can submit plain text and the service embeds it for you.

Vector, hybrid and semantic search

flowchart TB
  Q["Query string"] --> A["Full-text search<br/>BM25 over the inverted index"]
  Q --> B["Vector search<br/>HNSW / exhaustive KNN<br/>over embeddings"]
  A --> RRF["Reciprocal Rank Fusion<br/>merges the two ranked lists"]
  B --> RRF
  RRF --> SEM["Semantic ranker (L2)<br/>Bing-derived models rerank,<br/>plus captions and answers"]
  SEM --> TOP["Top-k grounding chunks"]
  RRF -.->|"semantic ranking off"| TOP

from azure.search.documents.models import VectorizableTextQuery

# Hybrid + semantic: the benchmark-backed default for RAG grounding.
# search_text drives BM25, vector_queries drives ANN, RRF merges them,
# then the semantic ranker reorders the merged list.
results = search_client.search(
    search_text="what is the parental leave allowance for contractors",
    vector_queries=[
        VectorizableTextQuery(            # the index vectorizer embeds this for you
            text="what is the parental leave allowance for contractors",
            k_nearest_neighbors=50,
            fields="content_vector",
        )
    ],
    query_type="semantic",
    semantic_configuration_name="default",
    query_caption="extractive",
    query_answer="extractive",
    filter="department eq 'HR' and effective_date ge 2026-01-01",
    select=["chunk_id", "title", "content", "url"],
    top=5,
)

for r in results:
    print(r["@search.reranker_score"], r["title"])

Agentic retrieval, knowledge sources and knowledge bases

Agentic retrieval is a multi-query pipeline for complex agent workflows. A knowledge base represents a complete domain and consists of one or more knowledge sources, an optional LLM for query planning and answer synthesis, and retrieval parameters. Each request goes through planning, decomposition into focused subqueries, parallel retrieval, semantic reranking and merging, returning a three-part response optimised for agent consumption.

• Indexed knowledge sources use the classic indexing and query engines; remote knowledge sources bypass indexing and are queried live.
• Your agent references the knowledge base for what to ground on; the knowledge base decides how to ground.
• Foundry surfaces this as Foundry IQ — a managed, permission-aware knowledge layer reusable across agents.
• A query planner breaks conversational history into subqueries; the merged documents are the grounding data the model cites.

Connecting retrieval to agents and workflows

5.2 Extract content from documents

Content Understanding: the multimodal extraction pipeline

Content Understanding combines OCR, layout analysis and generative field extraction in one service, and returns confidence scores and source grounding so a workflow can decide what to accept automatically and what to send for review. Its stages:

# An analyzer is a reusable definition: what to extract, how, and in what shape.
analyzer = {
    "baseAnalyzerId": "prebuilt-document",
    "config": {
        "enableOcr": True,
        "enableLayout": True,               # tables, sections, reading order
        "enableFigureDescription": True,    # describe charts and diagrams
        "estimateFieldSourceAndConfidence": True,  # confidence + grounding
        "tableFormat": "html",
    },
    "fieldSchema": {
        "fields": {
            "InvoiceNumber": {"type": "string", "method": "extract"},
            "InvoiceDate":   {"type": "date",   "method": "extract"},
            "VendorName":    {"type": "string", "method": "extract"},
            "Total":         {"type": "number", "method": "extract"},
            "Category": {
                "type": "string",
                "method": "classify",
                "enum": ["goods", "services", "travel", "other"],
            },
            "RiskSummary": {
                "type": "string",
                "method": "generate",
                "description": "One sentence on anything unusual about this invoice.",
            },
        }
    },
}

Classification and splitting

A single API call can classify, split and extract. The supported document scenarios are worth memorising because they map directly onto exam wording:

1. Classify only — the file contains one document type; label it.
2. Classify and analyze — label it, then route it to the right extraction analyzer.
3. Classify and segment — one file holds several concatenated documents (a loan pack: application form, pay slip, bank statement); split it.
4. Classify, segment and analyze — split, then route each segment to its own analyzer.
5. Hierarchical classifier — a category can route to another classifier for a second level.

Classification needs no training dataset: you define up to 200 category names and descriptions in the analyze operation. From the GA version you must include an other category, or every file is forced into one of your defined categories. The minimum classification unit for documents is a single page — intra-page classification is not supported.

Clean, grounded representations for RAG

"Produce clean, grounded representations to use with agents and RAG" is the objective that explains why Content Understanding sits in this skill area at all. A raw PDF is a bad input to a vector store; a Content Understanding markdown output is a good one, because it:

• Preserves structure — headings, sections, tables (including tables spanning multiple pages as one logical table), reading order.
• Captures figures — enableFigureDescription and enableFigureAnalysis turn charts and diagrams into text that can be indexed and retrieved. Information that is otherwise invisible to search.
• Detects annotations — handwritten notes, underlines, strikethroughs.
• Normalises every modality into the same shape, so audio and video land in the same index as documents.
• Carries grounding and confidence, so citations point at a real region of a real page.

Content Understanding vs Document Intelligence