Microsoft Azure AI Apps and Agents Developer Associate

Question 1

You maintain an Azure AI Search enrichment pipeline that ingests 200,000 scanned PDFs. The skillset runs OCR, a custom Web API skill for entity extraction, and text splitting. During development you frequently modify only the text-splitting skill's parameters and re-run the indexer, but every run re-executes the expensive OCR and custom Web API skills against all documents, driving up costs and runtime. You need to reduce redundant billed skill executions across iterative runs while keeping unchanged upstream enrichments intact. What should you do?

Accepted Answer

Enable incremental enrichment by configuring an enrichment cache (cache storage connection) on the indexer definition so only changed skills and their downstream skills re-execute.

Answer

Set the indexer schedule to run less frequently and increase the batch size so fewer total runs occur.

Answer

Add a conditional skill before OCR that skips documents already present in the index based on their document key.

Answer

Switch the OCR and custom Web API skills to run in a separate skillset and merge the outputs with a knowledge store projection.

Question 2

You maintain an Azure AI Search index that grounds a RAG-based support agent. Documents are stored in an Azure Blob Storage container that receives roughly 50 updated or new files per day out of 2 million total blobs. The current indexer runs on a nightly schedule but reprocesses the entire container each run, consuming large amounts of enrichment (skillset) cost and taking hours. You need the indexer to process only blobs that have changed since the last successful run, with the least operational effort. What should you configure?

Accepted Answer

Rely on the indexer's default change detection for Azure Blob Storage, which uses the blob's LastModified timestamp as the high-water mark.

Answer

Enable soft delete on the storage account and add a deletion detection policy so the indexer skips unchanged blobs.

Answer

Add a custom Web API skill that compares document hashes and short-circuits enrichment for unchanged files.

Answer

Switch the indexer to a full reset and re-run on demand only when the source data owner notifies you of changes.

Question 3

You build an Azure AI Search skillset that runs OCR, key phrase extraction, and a custom entity skill over a large collection of scanned contracts. Beyond serving a searchable index, your data science team wants the enriched, structured output (extracted entities and key phrases per document) persisted to Azure Blob Storage as tables and objects so they can run separate downstream analytics in Azure Machine Learning without re-invoking the skillset. What should you configure in the indexer/skillset to meet this requirement?

Accepted Answer

A knowledge store with table and object projections that map enriched nodes to Azure Storage

Answer

An incremental enrichment cache pointing at the Blob container used by the analytics team

Answer

An index projection that splits parent documents into child chunks in the search index

Answer

An output field mapping that copies enriched values into additional index fields

Question 4

You are building an Azure AI Search index to support a RAG grounding pipeline over a large corpus of policy documents. Each document is chunked, and every chunk carries a text passage, a precomputed embedding vector, and a 'department' value used to scope queries. Requirements: users must be able to run hybrid (keyword + vector) queries, filter results by department, and see the original passage text returned in results, but the raw embedding vector should never be returned in the response payload to reduce bandwidth. How should you configure the field attributes in the index definition?

Accepted Answer

Make the vector field searchable but not retrievable, the passage field searchable and retrievable, and the department field filterable and retrievable.

Answer

Make the vector field searchable and retrievable, the passage field filterable, and the department field searchable.

Answer

Make the vector field filterable, the passage field retrievable only, and the department field searchable and retrievable.

Answer

Make all three fields searchable, retrievable, and filterable so the query engine can decide which to use at runtime.

Question 5

You are building an Azure AI Search ingestion pipeline for a repository of scanned PDF contracts. Each PDF contains both embedded digital text pages and image-only scanned pages. You must produce a single searchable text field per document that combines the native text with text recognized from the scanned images, so downstream RAG retrieval has complete grounding content. Which skillset configuration should you use in the indexer?

Accepted Answer

Add an OCR skill to extract text from images, then a Text Merge skill to combine the OCR output with the document's native content field into one consolidated text field

Answer

Add only the OCR skill and map its output directly to the index, relying on it to also capture the embedded digital text

Answer

Add a Document Extraction skill followed by an Entity Recognition skill and index the recognized entities as the grounding content

Answer

Add a Language Detection skill and a Key Phrase Extraction skill, then index the key phrases as the searchable grounding field

Question 6

A retail team uses an Azure OpenAI image-generation model to update product marketing images. They need to replace only the background of an existing photograph while keeping the product itself pixel-identical, and they must specify exactly which region should be regenerated. Which image-editing capability should they use?

Accepted Answer

Inpainting with a mask that marks the background region to be regenerated

Answer

Text-to-image generation using a detailed prompt describing the desired background

Answer

Image variations that produce alternative renderings of the whole source image

Answer

Outpainting to extend the canvas beyond the original image boundaries

Question 7

A legal-tech company processes thousands of lengthy contract documents overnight. They need each document condensed into a concise, fluent narrative overview that rephrases key clauses in natural language for busy executives, rather than pulling verbatim sentences. They want to use Azure AI Language's summarization capability. Which summarization approach should they configure?

Accepted Answer

Abstractive summarization, which generates new, rephrased sentences that capture the document's meaning

Answer

Extractive summarization, which ranks and returns the most relevant original sentences

Answer

Key phrase extraction, returning a list of the most salient noun phrases

Answer

Entity recognition, returning categorized named entities from each contract

Question 8

A logistics company uses an Azure AI Content Understanding custom analyzer to extract shipping fields (tracking number, weight, destination) from scanned bills of lading. Downstream, an agent auto-approves shipments based on the extracted values. Auditors flag that some low-quality scans produced incorrect destination values that were still auto-approved. You must ensure that uncertain extractions are routed to a human reviewer before the agent acts, without discarding well-extracted documents. What should you do?

Accepted Answer

Use the per-field confidence scores returned by the analyzer and route any document with a field confidence below a defined threshold to a human-in-the-loop review queue.

Answer

Switch the analyzer output format from structured JSON to markdown so the agent can read the raw layout text and infer correctness itself.

Answer

Increase the number of few-shot labeled examples in the analyzer schema so that all destination values are always extracted with 100% accuracy.

Answer

Enable OCR on the analyzer so that scanned documents are converted to text, which guarantees correct field extraction on low-quality images.

Question 9

You are building a RAG ingestion pipeline for a library of engineering white papers that contain descriptive charts and diagrams embedded within the text. Your grounding requirement is that when an agent answers a question about a chart, the retrieved chunk must include a natural-language description of what the chart depicts (trends, axes, labeled regions) rather than just the surrounding paragraph text, so the model can reason about the visual without receiving the raw image at query time. You are using Azure AI Content Understanding to produce the grounded representation. Which configuration produces the required output for downstream reasoning?

Accepted Answer

Configure a document analyzer that emits markdown output and enable generation of natural-language descriptions for embedded figures, so each diagram is represented as descriptive text inline with the surrounding content.

Answer

Use a layout-only analyzer that returns bounding-box coordinates and reading order for each figure region, then store the coordinates alongside the text chunks.

Answer

Run an OCR skill over the pages to extract any text printed inside the charts, and rely on that extracted text as the chart's grounded representation.

Answer

Extract each figure as a separate binary image projection and attach the raw image bytes to the chunk so the retriever can pass them to the model at query time.

Question 10

A legal-tech team uses Azure AI Content Understanding to extract clauses from scanned contracts and feed the results into a RAG agent. Downstream, compliance reviewers must be able to verify each extracted clause value by tracing it back to the exact location in the original document where it was found. Which capability of the Content Understanding analyzer output should the team rely on to satisfy this traceability requirement?

Accepted Answer

Enable the analyzer's field-level grounding (source spans / bounding regions) so each extracted value includes its location in the source document.

Answer

Increase the analyzer's confidence threshold so only high-confidence fields are returned.

Answer

Convert the analyzer output to markdown so the layout structure is preserved in the text.

Answer

Add a custom skill in Azure AI Search that re-runs OCR on the indexed chunks at query time.

Question 11

Your team is building a RAG pipeline for a legal firm. Source files are scanned contracts that combine paragraphs of prose, signature blocks, and embedded tables of payment schedules. Downstream, an agent must reason over both the narrative clauses and the tabular figures, and the retrieved chunks must preserve the reading order and table structure so the model does not confuse rows and columns. You are configuring an Azure AI Content Understanding analyzer to produce the grounded representation that will be chunked and indexed. Which analyzer output configuration best supports this requirement?

Accepted Answer

Configure the analyzer to emit Markdown output so tables are rendered with structured Markdown table syntax while prose is preserved in reading order.

Answer

Configure the analyzer to emit only a flat plain-text concatenation of all recognized words to maximize embedding density.

Answer

Configure the analyzer to emit a field-extraction schema that returns only the payment amounts as key-value pairs and discards the narrative text.

Answer

Configure the analyzer to emit raw OCR bounding-box coordinates per word and pass those coordinates directly to the embedding model.

Question 12

A logistics company processes three distinct document types (bills of lading, customs declarations, and delivery receipts) that arrive mixed together in a single intake queue. Each type has a different set of fields to extract, but all must be routed to the same downstream RAG grounding store as clean, structured output. You are building the ingestion stage with Azure AI Content Understanding. What is the most appropriate design to extract the correct fields per document type while minimizing misclassified extractions?

Accepted Answer

Create a separate custom analyzer with its own field schema for each document type, add a classification step to route each incoming document to the matching analyzer, then merge the structured outputs into the grounding store.

Answer

Build one analyzer whose field schema is the union of all fields across the three document types, and let it extract every field from every document regardless of type.

Answer

Use the prebuilt document analyzer to produce markdown for all documents and rely on the downstream LLM to infer document type and pull the correct fields at query time.

Answer

Configure a single analyzer with the highest confidence threshold so that only fields present in a given document are returned, eliminating the need to distinguish document types.

Question 13

A financial services company stores thousands of recorded customer support calls as audio files. They need a pipeline that produces, for each call, a structured output containing the full transcript, a concise summary, the detected primary topic, and the customer's stated account issue category (chosen from a fixed list of categories) so these fields can ground an Azure AI Foundry agent. They want to minimize custom code and avoid chaining multiple separate services manually. Which approach best meets these requirements?

Accepted Answer

Define an Azure AI Content Understanding audio analyzer with a field schema that specifies the transcript, summary, topic, and a classification field constrained to the category list, then run the audio files through the analyzer.

Answer

Send each audio file to the Speech-to-text batch transcription API, then write custom code to call a language model separately for summary, topic, and category extraction.

Answer

Index the raw audio files directly into Azure AI Search using integrated vectorization and rely on hybrid search to surface the summary and category at query time.

Answer

Use an Azure AI Vision analyzer with OCR enabled to read the audio waveform images and extract the required fields.

Question 14

A logistics company ingests a nightly batch of scanned PDFs into a blob container. Each PDF may be an invoice, a bill of lading, or a customs declaration, and the document type is not indicated in the file name or metadata. You need each document routed to the correct Azure AI Content Understanding analyzer so that the appropriate field schema is applied, without manually pre-sorting the files. What should you configure in the pipeline?

Accepted Answer

A Content Understanding classifier that categorizes each document and splits multi-type files, then routes each segment to the matching document analyzer

Answer

A single general document analyzer with a union of all three field schemas, letting downstream code ignore fields that are empty

Answer

An Azure AI Search indexer with an OCR skill and a shaper skill that infers the document type from extracted key phrases

Answer

A language service custom text classification model that reads the OCR output and assigns a category label before extraction

Question 15

A retail company wants to automate ingestion of supplier product photos. For each image, they need a structured JSON record containing the dominant color, whether packaging is present, and a short marketing description, using a schema they define once and apply across thousands of images. Which Azure capability should the developer use to define custom fields and extract these structured visual characteristics per image?

Accepted Answer

Azure AI Content Understanding with a custom analyzer schema defining the fields to extract

Answer

Azure AI Vision Image Analysis 4.0 with the standard tags and objects features

Answer

Azure OpenAI gpt-image image generation with a text prompt describing each product

Answer

Azure AI Content Safety image moderation categories

Question 16

A logistics company processes multi-page shipping manifests as scanned PDFs. Each page contains complex nested tables where cargo line items span multiple rows and are grouped under section headers. The data science team needs a clean, grounded representation that preserves the table structure (row/column relationships and section groupings) so a downstream RAG agent can reason accurately about which items belong to which shipment section. Which approach produces the most reliable structured representation for this requirement?

Accepted Answer

Use an Azure AI Content Understanding document analyzer configured to emit markdown output that preserves layout, table structure, and hierarchical headings for grounding.

Answer

Run the Read (OCR) model to extract raw text lines, then concatenate all detected text into a single plain-text blob for chunking.

Answer

Apply Azure AI Vision image captioning to each page to describe the tables in natural language before indexing.

Answer

Use key phrase extraction on the OCR output to identify the most important terms per page and index those as the grounded representation.

Question 17

Your team is building a RAG solution over engineering maintenance manuals stored as PDFs. Each manual mixes narrative paragraphs, exploded-view diagrams with callout numbers, and specification tables. The downstream agent must reason over the text AND correctly associate diagram callout numbers with their referenced part descriptions in the tables. You need to produce a clean, grounded representation that preserves the reading order, table structure, and figure-to-text relationships in a single pass. Which approach best meets these requirements?

Accepted Answer

Use a Content Understanding multimodal document analyzer that performs layout analysis, OCR, and figure understanding to emit structured markdown preserving reading order, tables, and figure captions/callouts.

Answer

Run the Azure AI Vision Read (OCR) API on each page to extract raw text lines, then concatenate the text in top-to-bottom order for chunking.

Answer

Build an Azure AI Search skillset with only the built-in OCR skill and merge skill, indexing the flat extracted text without layout or figure enrichment.

Answer

Use a text-only prompt to a language model asking it to describe each PDF page, relying on the model's general knowledge to reconstruct the tables and diagrams.

Question 18

A financial services team uses Azure AI Content Understanding to extract line items from vendor invoices. They need the extracted vendor names and account codes validated against an authoritative master list of approved vendors so that the analyzer can flag entries that do not match a known record. Which capability should they configure to achieve this validation during extraction?

Accepted Answer

Use Content Understanding pro mode with a reference data source (knowledge base) linked to the analyzer so extracted fields are cross-referenced against the approved vendor list

Answer

Increase the analyzer's confidence threshold so only high-confidence vendor extractions are returned

Answer

Add a custom OCR skill to the ingestion skillset to improve character recognition of vendor names

Answer

Enable markdown output on the analyzer so the vendor names appear as headings for easier manual review

Question 19

You are building a voice-enabled agent for a hospital's radiology department. During testing, the standard Azure AI Speech-to-text model frequently mis-transcribes specialized medical terms (e.g., 'oligodendroglioma', 'cholecystectomy') that clinicians dictate, even though general conversational speech transcribes accurately. You have a corpus of correctly spelled domain terms and representative sentences but only a small amount of matching audio. Which approach best improves recognition accuracy for these terms?

Accepted Answer

Train a custom speech model in Speech Studio using plain-text and structured-text (pronunciation) training data built from the domain corpus, then deploy it to a custom endpoint the agent calls.

Answer

Switch the audio format to a higher sample rate and increase the microphone gain to capture clearer input before sending it to the standard model.

Answer

Add a post-processing LLM prompt that rewrites the standard transcription output, instructing it to guess and correct any misspelled medical words.

Answer

Enable text-to-speech neural voice output so the agent reads the terms back to confirm them with the clinician.

Question 20

A support platform team is building an agent that reviews written customer chat transcripts. They need to detect whether messages express negative sentiment AND identify specific opinion targets (e.g., 'the checkout was slow' should link the negative opinion to 'checkout'). They want a managed Azure AI Language capability rather than crafting their own generative prompt, to keep results consistent and low-cost at high volume. Which capability should they use?

Accepted Answer

Sentiment analysis with opinion mining (aspect-based sentiment) in Azure AI Language

Answer

Key phrase extraction in Azure AI Language

Answer

Named entity recognition (NER) in Azure AI Language

Answer

A custom GPT-4o generative prompt returning a free-form sentiment paragraph

Question 21

A media company uses an Azure OpenAI image-generation deployment (gpt-image) to let marketing staff create promotional artwork from free-text prompts. Compliance requires that the pipeline automatically block any generated image containing sexual, violent, or hateful visual content before it is delivered to the end user, without the developer writing custom classifier code. Which approach best meets this requirement?

Accepted Answer

Rely on the built-in Azure OpenAI content filters, which evaluate the input prompt and the generated image and return a contentFilter error instead of an image when a category exceeds the configured severity threshold.

Answer

Send each generated image to Azure AI Vision Image Analysis and inspect the returned tags for words like 'violence' or 'nudity' to decide whether to block it.

Answer

Add a Prompt Shield policy to the deployment so that unsafe visual output is filtered by the jailbreak-detection model.

Answer

Enable Azure AI Content Understanding on the deployment to extract visual characteristics and reject any image whose caption references restricted concepts.

Question 22

A marketing team uses an image-generation model in Azure AI Foundry to produce promotional assets. Compliance requires that every published image include the company watermark and must never contain competitor logos or other prohibited brand symbols. The team wants an automated post-generation step that inspects each produced image, verifies the presence of the required watermark, and flags any prohibited brand marks before the asset is released. Which approach best enforces these visual policy rules?

Accepted Answer

Add a post-generation validation stage that uses Azure AI Vision brand and logo detection plus a watermark-presence check, rejecting images that fail either rule before release.

Answer

Lower the generation model's temperature so it deterministically reproduces the watermark and avoids drawing competitor logos.

Answer

Enable the Azure OpenAI content filter with a high severity threshold for the hate and violence categories to block prohibited brand usage.

Answer

Attach the required watermark text as a negative prompt so the model refuses to generate competitor logos.

Question 23

A product team collected 40,000 free-text customer feedback comments with no predefined categories. They want to discover the recurring themes across the corpus—without maintaining a fixed label list—so they can prioritize product improvements. Which Azure AI Language capability best fits this requirement?

Accepted Answer

Key phrase extraction to pull salient terms from each comment individually

Answer

Custom text classification with a labeled training set mapping comments to known categories

Answer

The abstractive summarization API run once over the entire concatenated corpus

Answer

Named entity recognition (NER) to tag product and organization entities in each comment

Question 24

You are building a semiautonomous customer-support agent in Azure AI Foundry. The agent can call a custom function that issues refunds directly to customer payment methods. Business rules require that any refund over $500 be reviewed and explicitly approved by a human support supervisor before execution, while smaller refunds may proceed automatically. You need to implement this with the least custom orchestration code while keeping an auditable record of decisions. What should you do?

Accepted Answer

Configure a human-in-the-loop approval flow that intercepts the refund tool call, pauses the run when the amount exceeds $500, and requires a supervisor to approve or reject before the tool executes.

Answer

Lower the model's temperature to 0 so the agent behaves deterministically and never issues incorrect refunds above the threshold.

Answer

Add a system instruction telling the agent to never issue refunds over $500 and rely on the model to enforce the limit.

Answer

Enable a content safety filter with a high severity threshold to block refund tool calls that exceed the dollar limit.

Question 25

You are building a customer-support agent in Azure AI Foundry Agent Service. The agent answers product questions from an internal knowledge base, but users also ask about breaking news that affects your products (e.g., recent regulatory changes and current market prices) that changes daily and is not present in your indexed documents. You must add a capability so the agent can retrieve up-to-date public web information at query time while keeping your existing knowledge-base retrieval for internal content. Which tool should you add to the agent?

Accepted Answer

Grounding with Bing Search tool

Answer

File Search tool pointing to a vector store of uploaded documents

Answer

Code Interpreter tool

Answer

Azure AI Search connected index tool

Question 26

You are building an Azure AI Foundry agent for a finance team. Users upload CSV files containing quarterly sales data and ask the agent to compute custom aggregations, generate correlation statistics, and produce chart images on demand. The requested calculations vary unpredictably per user and cannot be pre-defined as fixed functions. Which built-in agent tool should you enable to satisfy these requirements?

Accepted Answer

Code Interpreter tool, which runs generated Python in a sandboxed environment against uploaded files

Answer

File Search tool, which indexes the uploaded files and performs semantic retrieval over their contents

Answer

A custom OpenAPI function tool that calls an external statistics microservice

Answer

Bing Grounding tool, which augments answers with live web search results

Question 27

Your team is building a generative application in Azure AI Foundry that summarizes long legal contracts. Business stakeholders require the highest possible summarization quality and are willing to accept higher per-token cost and slower response times, since summaries run as scheduled overnight batch jobs rather than interactively. You must choose a deployment approach in Foundry that best fits this workload. Which approach should you select?

Accepted Answer

Deploy a large frontier reasoning-capable model on a standard (pay-as-you-go) deployment and run the summaries as an offline batch job.

Answer

Deploy a small language model (SLM) optimized for edge latency to reduce cost per summary.

Answer

Configure a model router to dynamically pick the cheapest model per request to minimize spend.

Answer

Deploy a multimodal vision model so contract images and text can both be processed.

Question 28

Your team deploys a customer-support agent in Azure AI Foundry that answers questions using an Azure AI Search index built from a product knowledge base. Users report that the agent frequently returns answers based on outdated product documentation, even though the underlying source documents in Blob Storage were updated two weeks ago. The index was configured with an indexer that ran once during initial setup. What is the most appropriate action to ensure the agent grounds its answers on current content while minimizing manual effort?

Accepted Answer

Configure a schedule on the Azure AI Search indexer so it periodically re-crawls the Blob Storage data source and refreshes the index.

Answer

Increase the agent model's temperature so it generates more up-to-date responses from its training data.

Answer

Switch the agent's retrieval method from vector search to keyword-only search to improve accuracy.

Answer

Recreate the entire index manually each time a document changes and redeploy the agent.

Question 29

Your team deploys an Azure AI Foundry agent that calls a third-party payment API requiring a static API key. Security policy mandates that the key never appears in code, configuration files, or the Foundry project settings in plaintext, and that the key can be rotated without redeploying the agent. Which approach best satisfies these requirements while integrating the credential into the agent's tool call?

Accepted Answer

Store the API key as a secret in Azure Key Vault, grant the agent's managed identity the Key Vault Secrets User role, and reference the secret at runtime through a Foundry connection.

Answer

Embed the API key in the agent's OpenAPI tool definition and mark the field as sensitive so Foundry masks it in the portal.

Answer

Add the API key as a plaintext environment variable in the managed online endpoint's deployment configuration.

Answer

Hardcode the API key in the agent's prompt template and enable content filtering to prevent it from being echoed back.

Question 30

You are building an Azure AI Foundry agent for a customer support scenario where conversations can span dozens of turns. Users complain that after long sessions the agent begins to lose track of earlier details, and you also notice per-request token costs climbing steeply as the thread grows. You want the agent to retain the essential context of the entire conversation while keeping the prompt token count bounded. What is the most appropriate approach?

Accepted Answer

Implement a conversation summarization strategy that periodically condenses older turns into a running summary while keeping recent turns verbatim in the thread.

Answer

Increase the model deployment's max token limit and disable any truncation so the full thread is always sent to the model.

Answer

Lower the temperature parameter to 0 so the model focuses more consistently on the earliest messages in the conversation.

Answer

Create a new thread after every few turns and discard the previous thread to reset the token count.

Question 31

You are building a customer support agent with the Azure AI Foundry Agent Service. Each customer has an ongoing support case that may span several separate sign-in sessions over multiple days. Within a single session, the agent must remember earlier turns, and when a customer returns days later you must resume the exact same case context without re-uploading history. You also need to inspect the individual tool calls and model steps that occurred during one specific request for debugging. Which combination of Agent Service constructs should you use to model this?

Accepted Answer

Persist one thread per customer case for cross-session continuity, add messages to that thread each session, and inspect a specific run's steps to debug the tool and model activity of a single request.

Answer

Create a new thread for every sign-in session and store the full transcript in the agent's system instructions so it is reloaded each time the agent is invoked.

Answer

Use a single run that stays open for the entire multi-day case and read run-level metadata to recover per-session conversation history.

Answer

Store all conversation history in a vector store attached via file search, creating a fresh thread and run for each message to keep state stateless.

Question 32

You operate a production Foundry agent that calls several external APIs through registered tools. Support reports intermittent failures where the agent returns incomplete answers. You need to diagnose which tool invocations are failing and correlate them with the specific user turns and model reasoning steps, without adding custom logging code to each tool. What should you do first?

Accepted Answer

Enable tracing in the Foundry project and inspect the captured spans for each run, filtering on tool-call spans that returned errors or exceptions

Answer

Increase the agent's temperature so it retries failed tool calls with more varied inputs

Answer

Reduce the max token limit on the model deployment to force shorter, more reliable responses

Answer

Switch the agent to a fully autonomous oversight mode so it can recover from failures without human intervention

Question 33

You build a customer-support agent in Azure AI Foundry that uses the File Search tool grounded on a vector store of product manuals. Compliance requires that every answer the agent returns must display which source document and passage supported each claim, so auditors can verify the response was grounded and not fabricated. What is the most appropriate way to satisfy this requirement?

Accepted Answer

Parse the citation annotations included in the message content returned by the run and render the referenced file names and quoted text to the user.

Answer

Increase the temperature parameter so the model surfaces more of its reasoning about which documents it consulted.

Answer

Enable the Code Interpreter tool to open each manual file and print the matching paragraphs into the response.

Answer

Add a system instruction telling the model to always type the document title it remembers using at the end of every answer.

Question 34

You are building an Azure AI Foundry agent that must answer employee questions using a set of 40 internal PDF policy documents that rarely change. You want the agent to retrieve relevant passages from these documents automatically during a conversation, without building and maintaining your own separate Azure AI Search index or writing custom retrieval code. Which approach should you use to give the agent access to these documents?

Accepted Answer

Configure the File Search tool on the agent and upload the PDFs to a managed vector store that the tool queries automatically.

Answer

Register a custom function tool that calls an external embeddings API and returns matching text for every user turn.

Answer

Paste the full text of all 40 PDFs into the agent's system instructions so the content is always in context.

Answer

Enable the Code Interpreter tool and instruct the agent to parse the uploaded PDFs on each request.

Question 35

You are building a semiautonomous Azure AI Foundry agent for an e-commerce operations team. The agent uses a custom function tool named submit_refund that calls an internal payments API to issue customer refunds. During load testing you notice that transient network timeouts occasionally cause the agent's runtime to retry the tool call, resulting in some customers being refunded twice. You must prevent duplicate refunds without disabling automatic retries. What is the most effective approach?

Accepted Answer

Add an idempotency key parameter to the submit_refund tool schema and have the payments API deduplicate requests that share the same key

Answer

Lower the agent's temperature parameter to 0 so the model produces the same tool arguments on every call

Answer

Set tool_choice to 'none' so the model stops invoking submit_refund automatically

Answer

Increase the function tool's max_tokens so the completion is never truncated mid-call

Question 36

You build a Foundry agent that calls a custom function tool named getShippingQuote. During a run, the external shipping API the function wraps is temporarily unavailable and the function raises an exception. Your current implementation lets the exception propagate, which terminates the entire run and returns a generic failure to the user. You want the agent to gracefully recover — informing the user and, when possible, continuing the conversation with an alternative — without crashing the run. What is the recommended way to handle this in your tool implementation?

Accepted Answer

Catch the exception inside the function, return a structured tool output containing an error field and a human-readable message, and submit it as the tool output so the model can reason about the failure.

Answer

Increase the agent's max_completion_tokens so the model has room to retry the failed tool call automatically.

Answer

Remove the getShippingQuote tool from the agent definition so runs never attempt to call an unreliable dependency.

Answer

Set the run's temperature to 0 so the model produces deterministic output when the tool fails.

Question 37

You are building a customer-support agent in Azure AI Foundry Agent Service for an insurance company. The agent must handle policy questions but must never provide legal advice, must always escalate coverage disputes to a human, and must respond only in the customer's language. During testing, the agent occasionally offers interpretations of legal liability and answers in English regardless of the customer's language. You need to correct this behavior with the least engineering effort while keeping the guidance consistent across all conversations. What should you do?

Accepted Answer

Refine the agent's system instructions (persona and goal definition) to explicitly state its role boundaries, prohibited topics, escalation rules, and language-matching requirement.

Answer

Lower the temperature parameter to 0 so the agent produces deterministic responses that avoid legal interpretations.

Answer

Add a content filter with a high severity threshold to block any output containing legal terminology.

Answer

Fine-tune the underlying model on a dataset of compliant support transcripts to remove legal-advice behavior.

Question 38

You are designing an Azure AI Foundry agent for a legal firm. The agent must answer questions grounded in a corpus of 50,000 contract PDFs that are updated weekly. Answers must cite the exact clause and document, and the firm requires that the retrieval approach handle both keyword-heavy legal terminology (e.g., specific statute numbers) and conceptual paraphrased questions. Which knowledge integration approach should you configure for the agent?

Accepted Answer

Connect the agent to an Azure AI Search knowledge source configured with a hybrid index (vector + keyword) and semantic ranking, with a scheduled indexer for weekly updates

Answer

Attach the PDFs directly to each agent thread as file uploads so the model reads them in context at query time

Answer

Fine-tune the base LLM on the contract corpus and redeploy the fine-tuned model weekly

Answer

Store all contract text in the agent's system prompt and refresh the prompt template weekly

Question 39

You build a Foundry agent that generates detailed compliance summaries. Users report that longer summaries are being cut off mid-sentence, and the response metadata shows a finish_reason of 'length'. The prompt and system instructions are well within the model's context window. Which change most directly resolves the truncated output?

Accepted Answer

Increase the max_tokens (max output tokens) parameter for the completion request

Answer

Lower the temperature parameter to make the output more deterministic

Answer

Reduce the top_p value to constrain token sampling diversity

Answer

Add a stop sequence to force the model to end responses cleanly

Question 40

You are building an Azure AI Foundry agent for a call-center analytics team. Agents must accept uploaded recorded customer support calls (audio files), produce accurate transcripts, then reason over the transcript to summarize the customer's issue and detect sentiment. The team wants the least custom orchestration code and prefers using a Foundry-managed capability to handle the speech-to-text step before passing text to the LLM. Which approach should you implement?

Accepted Answer

Register a custom function tool that calls Azure AI Speech to transcribe the audio, then let the agent's chat model summarize and analyze the returned transcript.

Answer

Attach the audio files to the agent's file search vector store so the model can retrieve and transcribe them at query time.

Answer

Send the raw audio bytes directly in the chat completion prompt to a text-only GPT deployment and instruct it to transcribe and summarize.

Answer

Use the code interpreter tool to write Python that plays the audio and manually types out the transcript for the model.

Question 41

Your team is building an Azure AI Foundry agent that must automatically process supplier invoices that arrive as scanned PDF and image files. The agent needs to extract structured fields (invoice number, line items, totals) from these documents, then reason over the extracted data to flag anomalies before routing them for payment. You need to choose the most appropriate Foundry capability for the document extraction step while keeping the reasoning inside the agent. Which approach should you adopt?

Accepted Answer

Integrate Azure AI Document Intelligence (prebuilt invoice model) as a tool the agent calls to extract structured fields, then pass the results to the LLM for anomaly reasoning

Answer

Send the raw PDF bytes directly to a text-only LLM deployment and prompt it to return the invoice fields as JSON

Answer

Use Azure AI Search with a vector index of the scanned images to retrieve invoice fields for the agent

Answer

Deploy a small language model at the edge to perform OCR and field extraction locally before the agent runs

Question 42

Your team is building a customer-support agent in Azure AI Foundry that must accept photos of damaged products uploaded by customers and generate a written triage summary describing the visible defects. You need to deploy a model into your Foundry project and expose it to the agent so it can process both the image and an accompanying text description in a single request. Which approach correctly enables this capability?

Accepted Answer

Deploy a multimodal (vision-capable) model such as GPT-4o to a project deployment, then have the agent send messages that include both image and text content parts.

Answer

Deploy a text-only chat model and pre-process each image with an OCR connector, passing only the extracted text to the model.

Answer

Deploy an embeddings model and use it to convert the image into a vector that the agent forwards to the chat completion endpoint.

Answer

Deploy a small language model (SLM) at the edge because it natively handles image inputs with the lowest latency.

Question 43

You are building an Azure AI Foundry agent that must call your company's internal Orders REST API, which is protected by OAuth 2.0 client-credentials flow against Microsoft Entra ID. You register the API as an OpenAPI-specified tool for the agent. When configuring the tool, you must ensure the agent can authenticate to the API without embedding long-lived secrets in the tool definition and while respecting the API's token audience requirements. Which authentication configuration should you use for the OpenAPI tool?

Accepted Answer

Configure the OpenAPI tool with a Managed Identity authentication connection so the agent acquires Entra ID tokens scoped to the API's audience at call time.

Answer

Configure the OpenAPI tool with anonymous authentication and rely on the agent's network isolation to protect the API.

Answer

Configure the OpenAPI tool with an API key connection and store the client secret directly in the OpenAPI specification's securitySchemes section.

Answer

Configure the OpenAPI tool to pass the end user's chat session token directly to the Orders API on each request.

Question 44

A financial services team is deploying an Azure AI Foundry agent that can execute stock trades through a registered tool. Regulatory policy requires that any trade exceeding $10,000 be reviewed by a licensed broker before execution, while smaller trades and read-only balance inquiries may proceed automatically to maintain responsiveness. Which oversight configuration best satisfies these requirements?

Accepted Answer

Configure a conditional human-in-the-loop approval mode that intercepts tool calls only when the trade amount exceeds the $10,000 threshold, allowing other actions to run autonomously

Answer

Set the agent to fully autonomous mode and rely on the content safety filter to block high-value trades

Answer

Require human approval for every tool invocation the agent attempts, regardless of the action or amount

Answer

Disable the trading tool entirely and have the agent only return recommendations that brokers manually execute

Question 45

You are building an Azure AI Foundry agent for a travel-booking assistant. The agent has three registered tools: get_user_preferences (reads a profile store), search_flights (queries an airline API using the retrieved preferences), and check_weather (queries a public weather API for the destination). During a single turn the agent must produce a recommendation that uses flight results filtered by user preferences, while weather information is independent. You want to minimize end-to-end latency without producing incorrect results. How should you design the tool-calling behavior for this turn?

Accepted Answer

Invoke get_user_preferences first, then invoke search_flights and check_weather in parallel once preferences are available.

Answer

Invoke all three tools in parallel immediately, then discard any results that depend on data that arrived late.

Answer

Invoke all three tools strictly sequentially in registration order to guarantee deterministic output.

Answer

Invoke check_weather first because it has no dependencies, then run get_user_preferences and search_flights in parallel.

Question 46

You are building a generative application in Azure AI Foundry that must answer employee HR questions using an existing Azure AI Search index that already contains vectorized and semantically enriched HR policy documents. You want to consume the index from your Foundry-based application through a supported, managed integration rather than writing custom HTTP calls to the search service, and you want the connection to be reusable across projects in the hub. What should you do?

Accepted Answer

Create an Azure AI Search connection in the Foundry hub and reference it through the Foundry SDK connector to consume the index.

Answer

Embed the Azure AI Search admin key directly in the application code and query the REST endpoint with a custom client.

Answer

Copy the search documents into a local vector store file and attach it to each application session at runtime.

Answer

Recreate the index inside Foundry by re-uploading and re-embedding all HR documents through the SDK before querying.

Question 47

You are building a customer support agent in Azure AI Foundry that uses a large reasoning model. Users complain that responses feel slow because the full answer only appears after the model finishes generating, which can take several seconds for long replies. The backend already uses the Foundry Agent Service SDK. Business requirements state that perceived responsiveness must improve without changing the model or reducing answer quality. What is the most appropriate change to the way the agent consumes model output?

Accepted Answer

Enable streaming so partial tokens are delivered incrementally to the client as they are generated

Answer

Lower the max_tokens parameter so responses complete faster overall

Answer

Increase the temperature parameter to make the model produce output more quickly

Answer

Switch to a provisioned throughput deployment to eliminate all generation latency

Question 48

You are building a customer support agent in Azure AI Foundry Agent Service. Users interact with the agent across multiple sessions over several days, and the business requires that the agent recall prior conversation context (previous questions, order numbers, and stated preferences) within an ongoing support case without you having to manually re-send the entire history in each request. Which approach should you use to persist and manage conversational context across turns?

Accepted Answer

Create a persistent thread per support case and add messages to that thread so the Agent Service maintains conversation state across turns and sessions.

Answer

Increase the model's max token context window and resend the full transcript with every request to guarantee the model sees all prior turns.

Answer

Store each user message in an Azure AI Search vector index and rely on retrieval grounding to reconstruct the full conversation on every call.

Answer

Deploy a separate agent instance for each user so that each agent statelessly remembers only that user's data in the deployment configuration.

Question 49

You are building an agent with the Azure AI Foundry Agent Service SDK. Your code creates a thread, adds a user message to it, and then wants the agent to process the conversation and produce a response. After creating the run, you notice your code immediately reads the thread messages but finds no assistant reply. What is the correct approach to reliably retrieve the agent's response?

Accepted Answer

Poll the run status until it reaches a terminal state such as 'completed', then list the thread messages to read the newly added assistant message.

Answer

Add the user message a second time so the agent detects the change and generates a response synchronously.

Answer

Create a new thread for every turn because assistant messages are only returned on freshly created threads.

Answer

Delete the run and re-create the agent definition, since assistant replies are attached to the agent object rather than the thread.

Question 50

Your team deploys a customer-support agent in Azure AI Foundry that has access to three tools: a knowledge-base retrieval tool, an order-lookup API tool, and a refund-processing API tool. Compliance requires that the agent be able to autonomously answer questions and look up orders, but any refund action must never be executed by the agent without an explicit human decision. You must configure this while keeping all three tools registered to the agent. What is the most appropriate way to govern the agent's behavior?

Accepted Answer

Configure a human-in-the-loop approval requirement scoped to the refund-processing tool so that its invocation is paused for human authorization while the other tools run autonomously.

Answer

Remove the refund-processing tool from the agent and instruct users to call the refund API manually through a separate portal.

Answer

Lower the agent's temperature setting so it is less likely to choose the refund tool during a conversation.

Answer

Add a content safety filter that blocks any response containing the word 'refund' before it reaches the user.

Question 51

You are building an Azure AI Foundry agent that answers customer questions about order status. The agent has a single registered function tool, `get_order_status`, that queries an internal API. During testing, you notice that for a subset of queries like 'Where is my package?' the model responds with a conversational apology instead of calling the tool, so customers never receive real tracking data. You need to guarantee the model invokes `get_order_status` on every turn while keeping the deployed model and tool schema unchanged. Which configuration change should you make?

Accepted Answer

Set the tool_choice parameter to force the specific `get_order_status` function on each run.

Answer

Increase the temperature parameter so the model explores more tool-calling options.

Answer

Lower the max_tokens value to prevent the model from generating a conversational reply.

Answer

Remove the function description from the tool schema so the model relies only on the name.

Question 52

You are building a Foundry agent for a logistics company. The agent calls a custom function tool that queries a shipment tracking API. During testing, the agent occasionally invents delivery dates that were not present in the tool's response, presenting them confidently to users. You need to reduce these fabricated values while keeping the tool integration. Which approach most directly addresses the root cause?

Accepted Answer

Add explicit system instructions requiring the agent to only report field values returned by the tool and to state 'unknown' when a value is absent, then validate with a groundedness evaluator

Answer

Increase the temperature and top_p generation parameters so the model produces more varied responses grounded in the tool output

Answer

Switch the custom function tool to a code interpreter tool so the model can compute delivery dates itself

Answer

Remove the tool schema's optional fields so the model has fewer parameters to reason about

Question 53

You are defining a function tool schema for a Foundry agent that books meeting rooms. The tool accepts a 'building' parameter, but your organization only has three valid buildings (North, South, East). During testing, the model occasionally invents building names like 'West Tower' that do not exist, causing the function to fail. Without changing the underlying function code, what is the most effective way to prevent the model from passing invalid building values?

Accepted Answer

Define the 'building' parameter in the JSON schema with an 'enum' constraint listing North, South, and East

Answer

Set the tool_choice to 'required' so the model always calls the function

Answer

Lower the temperature parameter to 0 so the model becomes fully deterministic

Answer

Increase max_tokens so the model has room to reason about valid building names

Question 54

You are building an Azure AI Foundry agent for a logistics company. The agent must be able to call an existing internal REST API that returns real-time shipment tracking data. The API is documented with an OpenAPI 3.0 specification and is secured behind Microsoft Entra ID. You need the agent to invoke this API as a tool while authenticating without storing any secrets in the agent configuration. Which approach should you use?

Accepted Answer

Register the API as an OpenAPI specified tool on the agent and configure it to authenticate using the agent's managed identity via Microsoft Entra ID.

Answer

Wrap the API in an Azure Function and expose it to the agent as a code interpreter tool, embedding an API key in the function's app settings.

Answer

Add the API as a knowledge source in the agent's vector index so the agent can retrieve shipment data during grounding.

Answer

Configure the API as an OpenAPI tool and supply a long-lived Entra ID access token as a connection string in the agent's tool definition.

Question 55

Your team is building an Azure AI Foundry agent that must call several internal microservices exposed by a partner platform. The partner already runs a standardized server that publishes its available operations, tool definitions, and dynamically changing capabilities through a well-known protocol endpoint, so you want the agent to discover and invoke those tools at runtime without you hand-authoring and maintaining an OpenAPI spec for each service. Which tool integration approach best fits this requirement?

Accepted Answer

Register a Model Context Protocol (MCP) tool that points the agent at the partner's MCP server endpoint

Answer

Author a static OpenAPI 3.0 specification for each microservice and register them as OpenAPI tools

Answer

Use the built-in Code Interpreter tool to call the microservices with generated Python requests code

Answer

Define individual custom function tools with local JSON schemas for every partner operation

Question 56

Your team is building a customer-support solution in Azure AI Foundry that must orchestrate multiple specialized agents: one agent retrieves order history from an internal database, another drafts responses grounded in policy documents, and a supervisor agent decides which sub-agent to invoke and combines their outputs into a single reply. The team wants a managed capability that coordinates this multi-agent collaboration without building custom routing logic. Which Foundry capability should you use?

Accepted Answer

Use the Foundry Agent Service connected-agents (multi-agent orchestration) capability to let a main agent delegate tasks to sub-agents

Answer

Deploy each agent independently and write custom application code that calls each model endpoint sequentially and merges the results

Answer

Configure a single agent with a large system prompt that instructs it to perform database lookups and policy drafting in one pass

Answer

Use Azure AI Search indexers to chain the retrieval and drafting steps through skillsets

Question 57

You operate a production chat application backed by an Azure AI Foundry model deployment. A newer version of the base model has been released, and you must roll it out to real users while minimizing risk. Your requirements are: expose the new version to only a small percentage of live traffic initially, keep the existing version serving the majority of requests, and be able to shift traffic gradually or roll back instantly if quality metrics degrade. Which deployment approach best meets these requirements?

Accepted Answer

Create a second deployment for the new model version and use traffic allocation to route a small percentage of requests to it, adjusting the split over time.

Answer

Update the existing deployment in place to the new model version so all traffic immediately uses the latest model.

Answer

Deploy the new version to a separate Foundry project and ask users to opt in by switching endpoints manually.

Answer

Increase the provisioned throughput quota on the current deployment and enable autoscale to absorb the new version's load.

Question 58

A team at a financial services firm is building a Foundry-based application that must solve complex multi-step compliance calculations where the reasoning path matters as much as the final answer. Auditors need the model to break problems into intermediate logical steps before producing a conclusion. During evaluation, the team notices that a standard chat completion prompt often jumps to answers that are numerically wrong on multi-part questions. Which approach should the team adopt to most reliably improve the correctness of these multi-step reasoning tasks?

Accepted Answer

Select a reasoning-optimized model and use chain-of-thought prompting that instructs the model to work through intermediate steps before stating the final answer.

Answer

Lower the temperature parameter to 0 and increase the max tokens so the model produces longer, more detailed answers.

Answer

Enable the content safety jailbreak shield so unsafe reasoning paths are filtered out before the answer is returned.

Answer

Increase the top_p (nucleus sampling) value to allow the model to explore a wider range of candidate tokens for each answer.

Question 59

Your team develops a RAG-based chat application in Azure AI Foundry. You maintain separate development, staging, and production Foundry projects. The lead wants automated promotion of validated changes through these environments using Azure DevOps, with the same evaluation suite executed as a quality gate before any change reaches production. Which approach best integrates the Foundry project lifecycle into a CI/CD pipeline while enforcing this gate?

Accepted Answer

Store project, prompt, and configuration assets in a Git repository, and configure a pipeline that deploys to staging, runs the Foundry evaluation suite as a gate, and only promotes to production if evaluation metrics meet defined thresholds

Answer

Manually export the project assets from development and re-import them into production through the Foundry portal after a developer reviews the changes

Answer

Enable autoscale on the production endpoint so that new model versions are automatically rolled out as traffic increases

Answer

Use a single shared Foundry project for all environments and rely on deployment slots to separate development and production traffic

Question 60

Your team manages an Azure AI Foundry hub that serves three separate projects: a customer-facing chatbot, an internal HR assistant, and a finance analytics tool. The HR assistant must connect to an Azure AI Search index containing sensitive employee records. Compliance requires that this index be accessible only to the HR assistant project and never exposed to the other two projects, while other resources (such as the shared Azure OpenAI model deployment) remain available to all three. How should you configure the Azure AI Search connection?

Accepted Answer

Create the Azure AI Search connection scoped to the individual HR assistant project rather than sharing it at the hub level.

Answer

Create the Azure AI Search connection at the hub level so all projects inherit it, then use RBAC on the search service to block the other two projects.

Answer

Create three identical hub-level connections and assign each a different name so only the HR project uses the correct one.

Answer

Store the search index key in the hub's shared key vault and distribute it only to HR assistant developers.

Question 61

You are deploying a customer-facing generative AI assistant on an Azure AI Foundry model deployment. The compliance team requires that any output containing self-harm content be blocked entirely, but they also report that legitimate medical-support responses are being incorrectly rejected because the default content filter is too aggressive on the 'violence' category. You must reduce false positives on violence while keeping the strictest possible blocking for self-harm. What should you do?

Accepted Answer

Create a custom content filter configuration on the deployment, set the self-harm severity threshold to block at low and above, and raise the violence threshold to block only at high severity.

Answer

Disable the built-in content filter entirely and implement all moderation logic in application code using regular expressions.

Answer

Switch the deployment from a Standard to a Provisioned throughput unit deployment so that content filtering severity can be tuned per request.

Answer

Enable Prompt Shields for jailbreak detection, which will automatically relax the violence category while tightening the self-harm category.

Question 62

Your team deploys a customer-facing chatbot built on an Azure AI Foundry model. During a security review, testers demonstrate that they can bypass instructions by embedding malicious commands inside user-supplied documents that the agent summarizes, causing it to leak its system prompt. You must add a specific guardrail that detects and blocks these indirect injection attempts embedded in external content, while continuing to allow legitimate document summarization. Which Azure AI Content Safety capability should you enable?

Accepted Answer

Prompt Shields with indirect attack (document) detection

Answer

Protected material text detection

Answer

The severity-based hate and violence text moderation category

Answer

Groundedness detection in the evaluation pipeline

Question 63

Your team runs a customer-support agent in Azure AI Foundry backed by a GPT-4o standard deployment. Finance reports that monthly spend has tripled unexpectedly, and leadership wants proactive notification before the budget is exceeded next month, plus visibility into which application is driving token consumption. Which combination of actions best meets these requirements with the least custom development?

Accepted Answer

Create an Azure Cost Management budget with alert thresholds on the resource, and enable diagnostic/metrics logging to Azure Monitor to break down token usage per deployment

Answer

Switch the deployment to provisioned throughput units (PTUs) so costs become fixed and no monitoring is needed

Answer

Lower the max_tokens parameter on every request and rely on the Foundry portal quota page to track spending

Answer

Enable Microsoft Defender for Cloud on the subscription to receive cost anomaly alerts and usage attribution

Question 64

You operate a customer-support classification model deployed through Azure AI Foundry. Over the past two months, support agents report that the model increasingly misroutes tickets, even though it performed well at launch and the model itself has not been retrained or changed. You suspect the nature of incoming ticket text has shifted due to a new product line. Which monitoring capability should you configure to confirm and quantify this problem?

Accepted Answer

Enable data drift monitoring to compare the statistical distribution of live production inputs against the baseline training/reference dataset

Answer

Configure a content safety filter with a lower severity threshold to catch problematic ticket text

Answer

Increase the provisioned throughput units (PTUs) to reduce latency during peak ticket volume

Answer

Enable trace logging to capture the full provenance of each agent tool invocation

Exam domain	Exam weight	Practice
Plan and manage an Azure AI solution	28%	Practice this topic
Implement generative AI and agentic solutions	33%	Practice this topic
Implement computer vision solutions	13%	Practice this topic
Implement text analysis solutions	13%	Practice this topic
Implement information extraction solutions	13%	Practice this topic

Microsoft Azure AI Apps and Agents Developer Associate (AI-103) practice exam & study guide

Start studying AI-103

All AI-103 study resources

AI-103 exam domains

Sample AI-103 questions

Key AI-103 terms

AI-103 frequently asked questions

Plan and manage an Azure AI solution (28%)

Implement generative AI and agentic solutions (33%)

Implement computer vision solutions (13%)

Implement text analysis solutions (13%)

Implement information extraction solutions (13%)