AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

AI-103 · LESSON 5

Implement information extraction solutions10–15% · 2 topics

Learn to build retrieval and grounding pipelines and to extract content from documents.

5.1Build retrieval and grounding pipelines

Concept

Grounding ties a model’s answers to verifiable information sources. In Azure AI Search, data flows data source → indexer → skillset (chunking, embedding creation, OCR) → index, and hybrid search with the semantic ranker retrieves relevant information. It can also serve as agent knowledge.

How it works / steps

  1. Decide the data source (Blob, SharePoint and so on) and the index schema
  2. Configure integrated vectorization (chunking and embedding creation)
  3. Enable hybrid search (keyword + vector) and the semantic ranker
  4. Put the retrieved context and sources in the prompt and have it answer with citations
  5. Use document-level access control so only documents the user may see are searched

Workplace example

For internal policy Q&A, ingest SharePoint documents into AI Search, filter by user permissions and cite the policy chapter number as the source in answers.

Common confusions and exam tips

Mini quiz

Q1. In internal document search, you want to handle both exact matches like “warranty period for model AB-1234” and paraphrases like “what to do when it breaks.” Which search method is appropriate?

  1. Vector search (embedding similarity) only
  2. Semantic reordering with the semantic ranker only
  3. Keyword search (BM25) only
  4. Hybrid search (keyword search + vector search)
Show answer and explanation

Answer: D Hybrid search (keyword search + vector search)

Hybrid search captures both proper-noun matches and semantic similarity. [Other options] Vector search alone is weak at exact model-number matches, and keyword search alone is weak at paraphrases. The semantic ranker reorders results and is not a search method itself.

Q2. You want to reorder search results so that documents more relevant to the question come to the top. Which capability is appropriate?

  1. Scoring profiles
  2. Semantic ranking
  3. Vector profile settings
  4. Query facets
Show answer and explanation

Answer: B Semantic ranking

Semantic ranking re-ranks based on semantic relevance. [Other options] Vector profiles configure embedding search. Scoring profiles boost by field weights or freshness, not semantic relevance. Facets are for filtering.

Reference links

5.2Extract content from documents

Concept

For document extraction there is Azure Content Understanding (define a schema and extract fields using generative AI) and Document Intelligence (prebuilt models for invoices, receipts, IDs and so on, layout analysis, custom extraction models). Extracted structured data feeds business systems and search indexes.

How it works / steps

  1. Decide the document types and the fields to extract
  2. Use prebuilt models for common documents such as invoices, and custom models or Content Understanding for proprietary formats
  3. Convert tables, paragraphs and headings to Markdown and similar with layout analysis and use it for RAG chunking
  4. Have people check fields with low confidence scores
  5. Apply validation rules before registering extraction results in systems

Workplace example

From purchase orders whose format differs by supplier, extract order number, items and quantity with Content Understanding, and have staff check only fields with confidence below 0.8.

Common confusions and exam tips

Mini quiz

Q1. You want to use scanned paper invoices in PDF as RAG grounding. What processing is needed at ingestion?

  1. Vectorize the PDFs as they are and register them in the index
  2. Extract only the PDF metadata and index it
  3. Store the PDFs as images without converting them to Markdown
  4. Convert them to text with an ingestion flow that includes OCR
Show answer and explanation

Answer: D Convert them to text with an ingestion flow that includes OCR

Text in scanned images cannot be searched unless OCR converts it to text. [Other options] Scanned PDFs are images, so text embeddings cannot be made without OCR. Metadata only or stored images do not make the body searchable.

Q2. You want to extract supplier name, amount and payment due date from invoices and pass them to matching. Which method is appropriate?

  1. A pipeline combining OCR, layout analysis and field extraction
  2. Vectorize invoices and reuse values from similar past invoices
  3. Read the full text with OCR and pull out number-like amounts with regular expressions
  4. Summarize invoice images with a multimodal model and pass the summary to matching
Show answer and explanation

Answer: A A pipeline combining OCR, layout analysis and field extraction

Layout-aware field extraction turns tables and items into structured data. [Other options] Summaries are not in the format matching needs. Regular expressions are fragile across formats. Reusing past invoice values is wrong.

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)