LEARN. PRACTICE. GROW.Concept and mechanism
Azure AI Search connects ingestion, indexing, and querying. An indexer reads a source; a skillset can enrich content. Source field mappings and enrichment output field mappings serve different purposes. A globally successful job does not establish that every field is populated or current. Propagating blob deletions requires suitable policy and conditions. Hybrid search combines text and vector signals; semantic ranker reorders candidates and cannot alone fix a document missing from initial retrieval. Control document access before using results as context. Record source version, date, and evidence so support can distinguish ingestion, retrieval, and generation defects.
Guided application
For document extraction, define fields, business validation, and review of uncertain results. In Document Intelligence v4.0 composition, API 2024-11-30, an explicit classifier routes documents to extractors; do not assume the implicit mechanism of earlier versions. A 202 response starts asynchronous tracking: retain the operation reference and wait for a terminal state before consuming fields. Content Understanding supports evaluating structured extraction across modalities against a desired schema. In fictional reconciliation, compare extracted fields with evidence and business rules before accepting amounts. High confidence does not authorize a payment. Use known documents to trace each stage and locate losses, empty mappings, and old content before proposing retraining.
The field exists in the enrichment tree but is empty in the index: check the output mapping and value path before changing the model.
Common pitfalls
Green job as completeness; reranking as complete retrieval; 202 as final result; extraction confidence as business approval.
Related topics: Services, identity, and operations · Generation, grounding, and evaluation · Agents, tools, and control
Demonstrate the source, transformation, index, retrieval, and decision path using known examples.
Reference: Indexer source enrichment and output mappings · AI-102 historical skills measured 2025-12-23; exam retired 2026-06-30