Implement information extraction solutions
Retrieval and Grounding Pipelines with Azure AI Search
CoreIngest and index documents, images, audio, and video into Azure AI Search; configure keyword, vector, hybrid, and semantic retrieval; enrich content with built-in and custom skills; build RAG ingestion with chunking and OCR; and connect the resulting retrieval to workflows and agent tools.
Aligned to the AI-103 study guide with skills measured as of April 16, 2026; guide and product behavior verified October 12, 2026.
Why this matters
Answer quality in RAG and agents is capped by retrieval quality. A pipeline that skips OCR, embeds whole documents without chunking, uses mismatched embedding models, or never refreshes its index will ground answers in the wrong text. AI-103 tests the indexing and query choices that prevent those failures.
Must Know
- An indexer pulls content from a supported data source such as Blob Storage, cracks documents, runs an optional skillset, and writes to the index through field and output field mappings. Schedules and change detection keep the index current.
- Skillsets add AI enrichment: built-in skills include OCR, image analysis, entity recognition, key phrases, language detection, translation, Text Split, the Document Layout skill, the Content Understanding skill, embedding skills, and the GenAI Prompt skill; custom skills call your own web API.
- The OCR skill reads text from normalized images, which the indexer generates when imageAction is set; the Text Merge skill then combines OCR text with the document's own text.
- Integrated vectorization chunks content (Text Split, Content Understanding, or Document Layout skills) and embeds it during indexing, and a vectorizer on the vector field embeds query text at search time. The vectorizer must use the same embedding model as indexing.
- Hybrid search runs full-text (BM25) and vector queries in one request and merges results with Reciprocal Rank Fusion (RRF). Semantic ranker then reranks the top 50 results with a language understanding model and can return captions and answers.
- Vector fields are not filterable themselves, so filters apply to companion text or numeric fields.
- Content Understanding's prebuilt-documentSearch, prebuilt-imageSearch, prebuilt-audioSearch, and prebuilt-videoSearch analyzers produce RAG-ready markdown, descriptions, and transcripts that can be indexed alongside text.
- Agents connect to retrieval through the Azure AI Search tool (one index, with query types such as vector_semantic_hybrid), a Foundry IQ knowledge base, or a toolbox, and workflows can call the same knowledge base or index as a step.
Compare and Distinguish
- Keyword versus vector versus hybrid search: exact term matching versus semantic similarity versus both merged with RRF.
- Hybrid search versus semantic ranker: retrieval and fusion versus a second-stage rerank of the top results.
- OCR skill versus Document Layout skill versus Content Understanding skill: text from images versus structure-aware markdown and chunks versus layout, figures, and chunking from Content Understanding.
- Built-in skill versus custom skill: managed enrichment versus your own web API in the pipeline.
- Indexer pull model versus push API: scheduled ingestion from supported sources versus your code sending documents directly.
Scenario examples
- Scenario: Scanned PDFs return no search hits for words visible on the page. Think: enable image extraction, add the OCR skill, and merge OCR text before chunking.
- Scenario: Users search with synonyms and paraphrases while compliance users search exact clause numbers. Think: hybrid search with semantic ranking.
- Scenario: Product manuals should be chunked by headings and embedded without custom code. Think: integrated vectorization with a layout-aware chunking skill and an embedding skill, plus a matching vectorizer.
Exam traps
- A vector-only index misses exact identifiers such as part numbers that keyword search finds.
- Semantic ranker reranks only the top 50 results; it does not search the whole corpus.
- Changing the embedding model requires re-embedding the content; a mismatched vectorizer degrades results.
- Enriched content does not reach the index unless output field mappings or index projections route it there.
- OCR does nothing unless the indexer is configured to extract normalized images.
Key takeaways
- Extract, chunk, enrich, and embed during indexing so that queries retrieve focused passages.
- Use hybrid search with semantic ranking as the default retrieval for grounding.
- Connect retrieval to agents through the Azure AI Search tool or Foundry IQ.
How it works
- During indexing, document cracking extracts text and images, skills enrich them in an enrichment tree, and mappings write the results into index fields.
- At query time, text and vector queries run in parallel, RRF fuses rankings, and semantic ranker optionally reorders the top results.
Objects and administrative surfaces
- Azure AI Search data sources, indexers, skillsets, indexes, and vectorizers.
- Import data (new) wizard for chunking and vectorization.
- Semantic configurations with title, keyword, and content fields.
- Foundry agent Azure AI Search tool and Foundry IQ knowledge bases.
When to use it
- Use integrated vectorization when you want Azure AI Search to manage chunking and embedding.
- Use the push API when content comes from a source that indexers do not support.
Security and governance implications
- Use managed identities for indexers and skills to reach data and Foundry resources without keys.
- Use document-level permissions or permission-aware knowledge sources when users may see different content.
Troubleshooting signals
- Check indexer execution history for documents that failed cracking or skill execution.
- Use the debug option in queries to see keyword, vector, and semantic subscores when relevance looks wrong.
More detail
- Ingest and index documents, images, audio, and video.
- Configure semantic, hybrid, and vector search for grounding.
- Enrich content with built-in and custom skills for text, images, and layout.
- Build RAG ingestion with OCR and connect retrieval to workflows and agent tools.
Ready for the quiz?
- What does RRF do in a hybrid query?
- Why must a vectorizer match the indexing embedding model?
- Which indexer and skill settings make text inside scanned images searchable?
- When would you add a custom skill to a skillset?
Related objectives
- D5.1.S1 — Ingest and index content, such as documents, images, audio, and video
- D5.1.S2 — Configure semantic search, hybrid search, and vector search for grounding
- D5.1.S3 — Implement enrichment by using custom or built-in skills for text, images, and layout
- D5.1.S4 — Configure RAG ingestion flow, including documents and using optical character recognition (OCR)
- D5.1.S5 — Connect retrieval pipelines directly to workflows and agent tools