4.2.2. Indexers, Skillsets, and Integrated Vectorization
💡 First Principle: The indexer pipeline automates the "ETL for embeddings" so you don't hand-write the chunk-and-embed loop. An indexer crawls a data source on a schedule, a skillset applies transformations (document cracking, chunking, embedding, enrichment), and integrated vectorization wires embedding directly into the pipeline at both index and query time.
Without automation, RAG ingestion is a fragile custom script monitoring sources, chunking, calling an embedding model, and pushing vectors. Azure AI Search replaces that with declarative components: indexers pull from 10+ Azure data sources (Blob, SQL, Cosmos DB, OneLake); skillsets chain AI transformations including the chunking skills from 4.1.2 and an embedding skill; integrated vectorization means you attach a vectorizer to the index so both stored chunks and incoming queries are embedded automatically with the same model. Incremental indexing keeps the index fresh as source data changes.
⚠️ Exam Trap: Writing custom code to monitor a data source, chunk, embed, and upsert is the anti-pattern the exam tests against — it's exactly what integrated vectorization and indexer/skillset pipelines exist to replace. A scenario emphasizing "keep the index fresh as documents change without custom plumbing" points to indexers with integrated vectorization, not a bespoke script.
Currency note (verified June 2026): The older Azure OpenAI "On Your Data" feature for wiring a Search index to a model is deprecated, with a firm retirement date of October 14, 2026. Microsoft directs new work to Foundry Agent Service with Foundry IQ (now GA as the dedicated knowledge/retrieval layer) for agentic retrieval. Direct Azure AI Search integration remains the durable pattern. Expect the exam to favor the Foundry-native grounding path.
Reflection Question: A team keeps a hand-written Python job to re-embed documents nightly. Which Azure AI Search capabilities replace that job, and what reliability benefit do they bring beyond saving code?
