AI-103 Questions 61-70: Retrieval Pipelines and Content Understanding

Agents and RAG apps are only as good as the content they can find. Implement information extraction solutions (10-15% of AI-103) covers the pipeline behind that: ingesting documents and images into Azure AI Search, enriching them with skills, chunking and vectorizing them, monitoring indexers, connecting retrieval to agents through Foundry IQ, and extracting clean, structured content with Azure Content Understanding.

This is the seventh post in our AI-103 series (Questions 61-70). Each question is an original practice scenario mapped to the official skills outline, not a copy of a real exam item.

As always, choose your answer before revealing it. Many questions here describe a symptom, so think about which stage of the pipeline is really failing.

Question 61: Text Inside Scanned Images Is Not Searchable

An indexer loads PDFs from Blob Storage into Azure AI Search. Many PDFs contain scanned forms and photos of signs, and words that appear only inside those images never match a search. What should the team configure?

A) Turn on semantic ranker for the index

B) Change the indexer’s parsing mode to JSON

C) Double the chunk size in the Text Split skill

D) Set the indexer’s imageAction to generateNormalizedImages, add the OCR skill, and use the Text Merge skill to combine the OCR text with the document text

👁 Reveal Answer

Correct Answer: D

Explanation: Azure AI Search extracts image content at indexing time, not at query time. Setting imageAction produces normalized images during document cracking, the OCR skill reads their text, and the Text Merge skill inserts that text back into the document so it ends up in a searchable field such as merged_content. Image extraction has an extra charge, and OCR beyond the free daily limit needs a billable Foundry resource attached. A reorders results but cannot find text that was never indexed. B is for JSON files. C changes chunk boundaries, not what is extracted.

Reference: Extract text and information from images

Question 62: Chunks That Follow the Document’s Headings

A company indexes long technical manuals with clear headings. Fixed-size chunks often cut a procedure in half or mix two unrelated sections, and answers lose context. The team wants chunks that follow the manual’s own structure, up to the third heading level. What should they use?

A) The Text Split skill with a smaller maximum page length

B) The Document Layout skill with markdown output and markdownHeaderDepth set to h3

C) The Key Phrase Extraction skill

D) The OCR skill on every page

👁 Reveal Answer

Correct Answer: B

Explanation: The Document Layout skill uses the Document Intelligence layout model to detect a document’s structure and output it as Markdown or text. In markdown mode, content is grouped by heading, and markdownHeaderDepth controls the deepest level; with h3, deeper headings such as h4 are rolled into their h3 section. A still cuts by length, just more often. C extracts phrases and does not chunk anything. D reads text from images but ignores structure. Note that this skill needs a billable Foundry resource beyond 20 documents per indexer per day.

Reference: Document Layout skill

Question 63: Vector Results Turn Into Nonsense

An index was built with integrated vectorization by using the AzureOpenAIEmbedding skill and text-embedding-3-small. Last week someone changed the index’s vectorizer to text-embedding-3-large to “improve quality.” Since then, vector queries return random-looking results. What is the fix?

A) Point the vectorizer back to the same embedding model the skill used, or re-index all content with the new model before switching

B) Turn on semantic ranker to fix the ordering

C) Increase k in the vector query

D) Add a scoring profile to the vector field

👁 Reveal Answer

Correct Answer: A

Explanation: With integrated vectorization, the embedding skill creates vectors at indexing time and a vectorizer on the index converts text queries into vectors at query time. Microsoft’s documentation states that the vectorizer must match the embedding model used to encode your content. Vectors from different models live in different spaces, so comparing them gives meaningless similarity scores. B, C, and D all rearrange or widen results that are already wrong.

Reference: Integrated vectorization in Azure AI Search

Question 64: Keeping Simple Lookups Fast and Cheap

An agent uses a Foundry IQ knowledge base. Most user questions are simple lookups such as “What is the Wi-Fi password policy?” The activity log shows an LLM planning step for every question, which adds latency and Azure OpenAI token costs. What should the team adjust?

A) Add more knowledge sources to the knowledge base

B) Raise the retrieval reasoning effort to medium

C) Set the retrieval reasoning effort to minimal, which skips LLM query planning and sends queries directly to the knowledge sources

D) Turn off semantic ranker on the search service

👁 Reveal Answer

Correct Answer: C

Explanation: In agentic retrieval, the low and medium retrieval reasoning efforts send the query and history to an LLM that plans focused subqueries. At minimal, that step is skipped and queries go straight to the knowledge sources, which reduces latency and LLM cost. Microsoft lists lower reasoning effort as a cost control. A adds more work per query. B adds even more planning. D removes the reranking that agentic retrieval relies on for each subquery.

Reference: Agentic retrieval in Azure AI Search

Question 65: “Success” but Documents Are Missing

An indexer’s last run shows a status of Success, but users report that about 40 recent documents cannot be found. What is the most likely explanation, and where should the team look?

A) A Success status means every document was indexed, so the problem must be in the app

B) Warnings always mean documents were skipped, so delete and recreate the indexer

C) The search service is out of storage, which is shown only in the billing page

D) A run can succeed while individual documents fail, as long as failures stay below the indexer’s max failed items setting; check the run’s execution history for errors and the failed item count

👁 Reveal Answer

Correct Answer: D

Explanation: Microsoft’s documentation states that an indexer run can be successful even if individual documents have errors, as long as the number of errors is below the max failed items setting. The execution history (up to the 50 most recent runs) shows each run’s errors, warnings, items processed, and items failed. Azure Monitor metrics can also track documents processed and skill invocations for alerting. A is the wrong assumption. B is wrong because warnings, such as an image with no text, are common and do not always mean a problem. C is not how storage limits are reported.

Reference: Monitor indexer status and results

Question 66: A Question With Several Parts

Employees ask an HR agent questions such as “How did the parental leave policy change between 2025 and 2026, and does it apply to contractors in Germany?” A single vector query usually retrieves only one part of the answer. What retrieval approach handles this best?

A) Use a Foundry IQ knowledge base with agentic retrieval, so an LLM breaks the question into focused subqueries that run in parallel and are semantically reranked before the results are merged

B) Increase top to 50 and send everything to the model

C) Upload the HR handbook to file search

D) Ask users to send one question at a time

👁 Reveal Answer

Correct Answer: A

Explanation: Agentic retrieval is a multi-query pipeline built for complex questions. With query planning, an LLM uses the question and chat history to generate focused subqueries (2025 policy, 2026 policy, contractor rules in Germany), runs them in parallel, reranks each set, and returns merged content with optional source references. Foundry IQ knowledge bases use this engine. B floods the model with loosely related text. C helps with uploaded files but does not plan subqueries across your sources. D shifts the work to users. Note that LLM query planning is a preview capability.

Reference: Agentic retrieval in Azure AI Search

Question 67: Reports Full of Charts

A RAG assistant answers questions from quarterly reports. Plain OCR captures the text, but the key numbers are inside charts and diagrams, so answers about trends are often wrong. The team wants clean Markdown for indexing that includes descriptions of figures and charts. Which Content Understanding analyzer fits best?

A) prebuilt-read

B) prebuilt-documentSearch

C) prebuilt-invoice

D) prebuilt-audioSearch

👁 Reveal Answer

Correct Answer: B

Explanation: prebuilt-documentSearch is one of Content Understanding’s RAG analyzers. It extracts paragraphs, tables, and figures, adds figure descriptions, represents charts and diagrams in text form (chart.js and mermaid.js), includes a short document summary, and produces chunked output ready for embedding. A is basic OCR without layout analysis or figure descriptions, which is what the team already has. C extracts invoice fields, not report content. D is for audio.

Reference: Content Understanding prebuilt analyzers

Question 68: Defining Fields in a Custom Analyzer

A legal team builds a custom Content Understanding analyzer for supplier contracts. It needs three fields: the termination notice period as written in the contract, a risk level of Low, Medium, or High, and a plain-language summary. Which method should the risk level field use?

A) Extract

B) Generate

C) Classify

D) A separate classifier API called after analysis

👁 Reveal Answer

Correct Answer: C

Explanation: Content Understanding fields use one of three methods. Extract pulls values exactly as they appear (documents only), which fits the notice period. Classify chooses from a predefined set of categories, which fits Low, Medium, or High. Generate creates values freely, which fits the summary. A suits the notice period, not a rating the contract never states word for word. B could produce any text, not one of the three allowed values. D is outdated, because the GA API moved classification into the analyzer itself and deprecated the dedicated classifier APIs.

Reference: What is Azure Content Understanding?

Question 69: Knowing When to Ask a Human

An accounts payable workflow extracts fields from supplier documents with Content Understanding. Most values are correct, but auditors want any uncertain value checked by a person, and reviewers want to see exactly where each value came from. What should the team enable?

A) A higher temperature on the analyzer’s model deployment

B) Confidence and grounding estimation on the analyzer, then route fields below a chosen confidence score to a review queue that shows the source region

C) Markdown output only, without fields

D) Video segmentation

👁 Reveal Answer

Correct Answer: B

Explanation: Content Understanding can return a confidence estimate from 0 to 1 for each field value and identify the specific regions in the content where each value was extracted or generated. Routing low-confidence values to people, with the source region highlighted, gives auditors a clear human-in-the-loop process. A adds randomness and is not an analyzer setting. C removes the structured fields the workflow needs. D applies to video and does not measure confidence.

Reference: What is Azure Content Understanding?

Question 70: Only Show What the User Is Allowed to See

An internal agent answers questions from SharePoint and Blob Storage content. Some documents are restricted to HR and legal staff. The agent must never use restricted content when answering other employees. What should the team implement?

A) Add “do not reveal HR or legal documents” to the agent instructions

B) Create a separate index and agent for every employee

C) Retrieve everything, then ask the model to remove anything the user should not see

D) Use a Foundry IQ knowledge base that synchronizes access control lists from the sources and runs queries under the caller’s Microsoft Entra identity, so permissions are enforced at query time

👁 Reveal Answer

Correct Answer: D

Explanation: Foundry IQ provides permission-aware responses. It can synchronize ACLs for supported sources, honor Microsoft Purview sensitivity labels, and run queries under the caller’s Entra identity, so only content the user is authorized to see is retrieved in the first place. A and C both rely on the model behaving correctly after restricted content is already in its context, which is not a security control. B does not scale and is hard to keep in sync with changing permissions.

Reference: What is Foundry IQ?

Study Tips for Information Extraction

• Follow the pipeline. Crack documents, then extract images and text, then enrich with skills, then chunk, then embed, then index, then query. Find which stage is failing.
• Images need help. Turn on imageAction, add OCR, and merge the text back in, or use Content Understanding.
• Structure-aware chunking comes from the Document Layout skill or Content Understanding; fixed-size chunks come from Text Split.
• Same model on both sides. The query vectorizer must use the same embedding model as the indexing skill.
• Content Understanding fields are Extract, Classify, or Generate, with optional confidence and grounding for review workflows.
• Security belongs in retrieval. Enforce permissions at query time, not in the prompt.

Keep Practicing for AI-103

Review text and speech in AI-103 Questions 51-60, or go back to Questions 1-10 to start a full review. The official AI-103 study guide lists every skill measured.