Domain 5 · Implement information extraction solutions
This domain is the supply side of everything in Domain 2. RAG is only as good as what it retrieves, and what it retrieves is only as good as what was ingested, enriched, and indexed. Studying it as the upstream half of RAG rather than as a separate topic makes both domains easier: when an agent gives a confidently wrong answer, the cause is usually somewhere in this pipeline.
Build retrieval and grounding pipelines
A pipeline starts by ingesting and indexing content, and the content is deliberately varied — documents, images, audio, and video, not text files alone.
Retrieval is then configured across three modes that are frequently compared. Vector search matches on semantic similarity, semantic search improves ranking of results, and hybrid search combines vector similarity with keyword matching so that exact terms — product codes, names, identifiers — are not lost to embedding drift. Hybrid is the usual recommendation when a scenario mentions both natural-language questions and precise identifiers.
Content is improved on the way in through enrichment with built-in or custom skills for text, images, and layout, and the RAG ingestion flow is configured end to end, OCR included, so that scanned material becomes searchable rather than opaque. The last step is connecting these retrieval pipelines directly to workflows and agent tools, which is the seam where this domain hands off to Domain 2.
Extract content from documents
Documents get their own treatment because their meaning is partly structural. Multimodal pipelines combine OCR, layout analysis, and field extraction — OCR recovers the characters, layout recovers the tables, headings, and reading order that tell you what those characters mean, and field extraction pulls the specific values you asked for.
The goal throughout is producing clean, grounded representations for agents and RAG using Content Understanding, with analyzers that emit structured or markdown outputs suitable for downstream reasoning. Markdown matters more than it first appears: preserving heading hierarchy and table structure in the retrieved chunk is often what lets a model answer a question about a document correctly.
Quiz · Domain 5
A RAG application must retrieve passages that are conceptually similar to a user's natural-language question even when no keywords overlap. Which Azure AI Search capability directly addresses this requirement?