Skip to content
Cheat sheets

AI-103 · Domain 5: Implement information extraction solutions · 13% of the exam

Task 5.1: Build retrieval and grounding pipelines

Ingesting and indexing documents, images, audio and video in Azure AI Search, keyword, semantic, hybrid and vector queries, enrichment skills, OCR in the ingestion flow, and connecting retrieval to workflows and agents.

Study it

Sample questions

Try each one before opening the answer. Every option is explained, with the Microsoft documentation page that proves it.

Question 1 · choose 1

Technicians search a parts catalog in Azure AI Search with queries such as "seal kit for P-4471 pump leaking at low pressure". Pure vector search finds similar descriptions but often misses the exact part number. Which query approach should grounding use?

  1. AA vector query with a category filter applied to the results
  2. BHybrid keyword and vector query, merged by RRF, with semantic ranking
  3. CA full-text query only, with fuzzy matching turned on for every term
  4. DA vector query with a higher k so that more neighbors come back
Show the answer and why

  • AA vector query with a category filter applied to the results

    Incorrect

    Filtering narrows the candidates; it doesn't add exact matching on the part number.

  • BHybrid keyword and vector query, merged by RRF, with semantic ranking

    Correct

    Hybrid search runs full-text and vector queries in parallel and merges them with Reciprocal Rank Fusion; keyword search catches exact codes, and semantic ranking improves the order.

  • CA full-text query only, with fuzzy matching turned on for every term

    Incorrect

    Fuzzy keyword search loses the conceptual matches that vectors find for the symptom part of the query.

  • DA vector query with a higher k so that more neighbors come back

    Incorrect

    More nearest neighbors are still chosen by similarity, so the exact part number can still be missed.

Product codes, jargon, dates and names favor keyword search; meaning favors vectors. A hybrid query gets both in one request, and the semantic ranker reorders the merged results.

Microsoft documentation

Lesson: EmbeddingsTask 5.1Report a problem with this question

Question 2 · choose 1

An index has a vector field filled by an indexer whose skillset embeds each chunk. Today the app calls an embedding model itself before every search. The team wants to send plain text and let the search service turn it into a vector. What should they add?

  1. AA scoring profile that boosts the vector field's weight in results
  2. BSemantic ranking turned on for every text query on the index
  3. CA vectorizer assigned to the vector field through a vector profile
  4. DA second embedding skill in the indexer's skillset for the queries
Show the answer and why

  • AA scoring profile that boosts the vector field's weight in results

    Incorrect

    Scoring profiles adjust relevance of text matches; they don't turn text into embeddings.

  • BSemantic ranking turned on for every text query on the index

    Incorrect

    Semantic ranking reorders text results; it doesn't create query vectors.

  • CA vectorizer assigned to the vector field through a vector profile

    Correct

    A vectorizer is defined in the index and calls an embedding model at query time, so plain-text queries work against vector fields.

  • DA second embedding skill in the indexer's skillset for the queries

    Incorrect

    Skills run during indexing. They don't vectorize queries.

Integrated vectorization has two halves: an embedding skill in the skillset vectorizes content during indexing, and a vectorizer in the index vectorizes queries. Use the same embedding model for both.

Microsoft documentation

Task 5.1Report a problem with this question

Question 3 · choose 1

An indexer loads PDFs from Blob Storage into an index for a RAG app. Many of the PDFs are scanned paper forms, and their content field comes back empty, while digital PDFs index fine. What should you configure?

  1. AAdd a vectorizer so the scanned pages can be searched by vector
  2. BReset the indexer so that every PDF is processed again
  3. CAdd a text split skill to cut the empty content into smaller chunks
  4. DSet imageAction to generateNormalizedImages and add the OCR skill
Show the answer and why

  • AAdd a vectorizer so the scanned pages can be searched by vector

    Incorrect

    A vectorizer embeds queries; it doesn't extract text from images.

  • BReset the indexer so that every PDF is processed again

    Incorrect

    Reprocessing with the same configuration produces the same empty content.

  • CAdd a text split skill to cut the empty content into smaller chunks

    Incorrect

    Splitting works on text that already exists; empty content stays empty.

  • DSet imageAction to generateNormalizedImages and add the OCR skill

    Correct

    Normalized images expose the scanned pages to the enrichment pipeline, and the OCR skill extracts printed and handwritten text from them.

Text in scanned pages is an image until OCR reads it. Turn on image extraction with imageAction, then add the OCR skill to the skillset.

Microsoft documentation

Task 5.1Report a problem with this question

Question 4 · choose 2

A Foundry agent grounds answers through the Azure AI Search tool. The project and the search service are moving into a private virtual network, and the security team requires keyless access. Which two actions are required? (Choose TWO.)

  1. AAuthenticate the search connection with a Microsoft Entra managed identity
  2. BAllow public network access on the search service from all networks
  3. CSwitch the tool's query_type to simple for every agent request
  4. DEnable RBAC on the search service and assign the identity its search roles
  5. EKeep the admin API key in Key Vault and reference it from the connection
Show the answer and why

  • AAuthenticate the search connection with a Microsoft Entra managed identity

    Correct

    With private virtual networking the Azure AI Search tool needs managed identity authentication; key-based authentication isn't supported.

  • BAllow public network access on the search service from all networks

    Incorrect

    Opening the service to all networks contradicts the move to a private network and doesn't change authentication.

  • CSwitch the tool's query_type to simple for every agent request

    Incorrect

    The query type changes how the index is searched, not how the agent authenticates.

  • DEnable RBAC on the search service and assign the identity its search roles

    Correct

    Keyless access needs role-based access enabled on the service, with Search Index Data Contributor and Search Service Contributor assigned to the identity.

  • EKeep the admin API key in Key Vault and reference it from the connection

    Incorrect

    That is still key-based authentication, which isn't supported with private virtual networking for this tool.

In a private network the agent reaches Azure AI Search with a managed identity, never a key. Turn on RBAC for the search service and give the identity the index data and service roles the tool needs.

Microsoft documentation

Task 5.1Report a problem with this question

Question 5 · choose 1

A team queries a Foundry IQ knowledge base from a lightweight web page with no model of its own. Today the knowledge base returns raw grounding chunks, but the page needs a natural-language answer with citations. What should the team enable?

  1. AData extraction mode with a higher number of chunks per query
  2. BSemantic answers on a classic search query
  3. CA scoring profile on the knowledge base's underlying search index
  4. DAnswer synthesis in the knowledge base, with its LLM
Show the answer and why

  • AData extraction mode with a higher number of chunks per query

    Incorrect

    Data extraction is the default that returns raw chunks.

  • BSemantic answers on a classic search query

    Incorrect

    Semantic answers are verbatim passages from the index, not generated answers.

  • CA scoring profile on the knowledge base's underlying search index

    Incorrect

    Scoring profiles reorder results; they don't write answers.

  • DAnswer synthesis in the knowledge base, with its LLM

    Correct

    Answer synthesis (preview) answers queries in natural language with citations, instead of the default data extraction.

Knowledge bases return extractive data by default so an agent can reason over it; answer synthesis lets the knowledge base compose the answer itself when the caller has no model.

Microsoft documentation

Task 5.1Report a problem with this question

Question 6 · choose 1

A documentation site uses classic Azure AI Search with semantic ranking. It wants a short, verbatim passage that directly answers the query shown above the result list, with no generative model involved. What should it request?

  1. ASemantic answers on the semantic query
  2. BA field mapping that copies the title into the content
  3. CAnswer synthesis from a knowledge base's LLM
  4. DA vector query with a higher k value
Show the answer and why

  • ASemantic answers on the semantic query

    Correct

    A semantic answer is verbatim content that a reading comprehension model recognized as answering the query; it isn't generated.

  • BA field mapping that copies the title into the content

    Incorrect

    Field mappings shape indexing, not answers at query time.

  • CAnswer synthesis from a knowledge base's LLM

    Incorrect

    Synthesis generates answers with an LLM, which the site wants to avoid.

  • DA vector query with a higher k value

    Incorrect

    More neighbors return more documents, not an answer passage.

Semantic ranking can add captions and answers extracted from the top documents; for composed answers, use RAG or answer synthesis.

Microsoft documentation

Task 5.1Report a problem with this question

Question 7 · choose 1

A blob indexer uses each blob's storage path as the search document key, and indexing fails because paths contain characters that aren't valid in keys. What should the indexer definition use?

  1. AA reset of the indexer before each run
  2. BA field mapping with base64Encode applied to the key
  3. CA Text Split skill that shortens each path before indexing
  4. DA field mapping with urlDecode applied to the key
Show the answer and why

  • AA reset of the indexer before each run

    Incorrect

    Resetting reprocesses documents with the same invalid keys.

  • BA field mapping with base64Encode applied to the key

    Correct

    base64Encode performs URL-safe encoding at indexing time, the documented fix when the source field for a key holds URL-unsafe characters.

  • CA Text Split skill that shortens each path before indexing

    Incorrect

    Skills work on content in the in-memory enrichment tree; the key comes from a source field through a field mapping.

  • DA field mapping with urlDecode applied to the key

    Incorrect

    urlDecode turns URL-encoded text back into plain characters, the opposite of making a key URL-safe.

Field mappings connect source fields to index fields and can transform values on the way, for example encoding paths into valid document keys.

Microsoft documentation

Task 5.1Report a problem with this question

Question 8 · choose 1

In a news archive's classic keyword search, older articles often outrank recent ones on the same topic. Editors want recency to boost results without changing the query text. What should the index define?

  1. AA scoring profile with freshness on the publish date
  2. BAn indexer schedule that runs every five minutes
  3. CA vector profile with a larger embedding model behind it
  4. DA field mapping that renames the publish date field in the index
Show the answer and why

  • AA scoring profile with freshness on the publish date

    Correct

    Scoring profile functions support freshness boosting on datetime fields.

  • BAn indexer schedule that runs every five minutes

    Incorrect

    More frequent indexing doesn't reorder results.

  • CA vector profile with a larger embedding model behind it

    Incorrect

    Embeddings capture meaning, not publication date.

  • DA field mapping that renames the publish date field in the index

    Incorrect

    Renaming doesn't change ranking.

Scoring profiles tune classic relevance with field weights and functions for freshness, distance, magnitude and tags.

Microsoft documentation

Task 5.1Report a problem with this question

Question 9 · choose 1

Each blob in a container holds a JSON array of 500 product records, and every record should become its own search document for retrieval. How should the blob indexer be configured?

  1. AAdd a Text Split skill to divide each blob into chunks
  2. BTurn on semantic ranking for the index
  3. CSet parsingMode to jsonArray for the blob indexer
  4. DLeave the default parsing so each blob is one document
Show the answer and why

  • AAdd a Text Split skill to divide each blob into chunks

    Incorrect

    Splitting by length doesn't follow record boundaries.

  • BTurn on semantic ranking for the index

    Incorrect

    Ranking doesn't change how documents are created.

  • CSet parsingMode to jsonArray for the blob indexer

    Correct

    jsonArray parsing produces a search document for each element of a JSON array in a blob.

  • DLeave the default parsing so each blob is one document

    Incorrect

    The default treats each blob as a single document.

Blob indexers offer JSON parsing modes (jsonArray, jsonLines) for one-to-many indexing; check how document keys are disambiguated for records from the same blob.

Microsoft documentation

Task 5.1Report a problem with this question

Practise domain 5 →Practise all domains →