AI-103 Hard Questions — 25 Difficult Questions

Question 1

An agent is designed to summarize meeting transcripts but keeps including internal meeting notes in the final summary. What is the most effective way to address this?

Accepted Answer

Update the system prompt to clarify which sections of the transcript to include.. The system prompt is the primary control surface for defining an agent's scope, tone, and what content it should include or exclude. When an agent incorrectly includes internal meeting notes, the root cause is usually ambiguous instructions about which transcript sections are in-scope, so clarifying the inclusion/exclusion rules in the system prompt directly fixes the behavior. This is a prompt engineering fix, not a model parameter fix, because the issue is about instruction-following scope rather than randomness or creativity.

Answer

Lower the temperature.

Answer

Switch to a different model.

Answer

Add a 'stop_sequence' to the API call.

Question 2

You are implementing a RAG pipeline where documents are chunked before embedding generation. You notice that semantic sentences are frequently split across chunk boundaries, resulting in degraded retrieval quality. Which chunking strategy should you implement to resolve this?

Accepted Answer

Sliding window chunking with token overlap. Sliding window chunking with token overlap is correct because it preserves semantic continuity by carrying a portion of the previous chunk's tokens into the next chunk. This overlap ensures that sentences or ideas spanning a boundary appear intact in at least one chunk, so the embedding captures the full semantic unit. It directly addresses the root cause of degraded retrieval: fragmented context at chunk edges.

Answer

Fixed-character chunking with zero overlap

Answer

Paragraph-only splitting with maximum size limits of 10,000 tokens

Answer

Random token slicing

Question 3

You are deploying a model that requires significant GPU resources. You want to ensure the deployment only scales when the average CPU load exceeds 70%. Which feature of Azure AI managed endpoints do you configure?

Accepted Answer

The 'Endpoints' deployment configuration 'autoscaling' settings.. Azure AI managed online endpoints support autoscaling rules configured on the deployment, where you define a metric (such as average CPU utilization) and a threshold (for example, 70%) that triggers scale-out or scale-in. This is configured in the deployment's 'autoscaling' settings under the endpoint configuration, allowing the deployment to add or remove instances based on load.

Answer

The 'Load Balancer' settings in the VNET.

Answer

Azure Front Door health probe settings.

Answer

The 'Compute' instance restart policy.

Question 4

You are using an Azure AI model that supports versioning. You want to perform an A/B test by routing 10% of traffic to a new version. Where do you configure this?

Accepted Answer

The 'Traffic' configuration in the Azure AI endpoint.. Azure AI endpoints expose a 'Traffic' configuration where you assign percentage weights to deployed model versions, enabling A/B testing by routing a fraction of requests to a new version. Setting 10% to the new version and 90% to the current one achieves the desired split. This is native to the endpoint and requires no external load balancer.

Answer

Azure App Service routing rules.

Answer

Azure Load Balancer backend pools.

Answer

Azure Traffic Manager DNS settings.

Question 5

You are optimizing an enterprise RAG application in Azure AI Foundry for high accuracy and minimal latency. Which THREE advanced retrieval techniques should you consider implementing? (Choose THREE)

Accepted Answer

Semantic ranking to re-rank top search results using deep learning models. Option A (Semantic ranking) is correct because Azure AI Search's semantic ranker applies deep-learning language models to re-rank the top BM25/vector results, promoting the most contextually relevant passages and directly improving RAG answer accuracy. Option C (Hybrid search combining full-text BM25 and vector embeddings) is correct because fusing keyword and dense-vector retrieval via Reciprocal Rank Fusion captures both exact term matches and semantic similarity, which raises recall and precision over either method alone. Option E (Query expansion and re-writing) is correct because enriching the user prompt with synonyms and related terms broadens the candidate set and aligns the query with the indexed content, improving retrieval quality for ambiguous or short queries. Option B does not belong because random sentence selection without indexing provides no relevance signal or deterministic retrieval, guaranteeing poor accuracy. Option D does not belong because disabling vector search and relying on random string matching discards semantic retrieval entirely, which destroys both accuracy and the ability to handle natural-language queries.

Answer

Using random sentence selection without indexing

Answer

Disabling vector search entirely to rely solely on random string matching

Question 6

You are troubleshooting a Prompt Flow and suspect the issue lies in the input processing. Which THREE tools or techniques can you use?

Accepted Answer

Setting breakpoints in the visual designer.. Option A is correct because Prompt Flow's visual designer supports setting breakpoints on individual nodes, letting you pause execution and inspect the exact input payload entering the input-processing step. Option B is correct because the Trace feature records the inputs and outputs of every node in a flow run, so you can pinpoint where the input processing diverges from expectations. Option C is correct because execution logs capture runtime diagnostics, errors, and intermediate values that reveal what happened during input processing. Option D is not appropriate because deleting and recreating the resource group is a destructive, unrelated action that would not diagnose input processing and could cause data loss. Option E is not appropriate because raising the model temperature only affects the LLM's sampling randomness and does not help inspect or validate the input-processing logic.

Answer

Deleting and recreating the entire resource group.

Answer

Increasing the model temperature to see if the error persists.

Question 7

You are troubleshooting a low confidence score on specific fields extracted by your Azure AI Document Intelligence custom template model. Which TWO actions can help improve the extraction confidence for those fields? (Choose TWO)

Accepted Answer

Ensure all training documents are of high visual quality and free from blur or heavy skew.. Option C is correct because custom template models learn field positions and visual patterns from the labeled samples, so blurry, skewed, or low-quality training documents degrade the model's ability to reliably locate and read those fields, producing low confidence scores; supplying sharp, straight, high-quality scans directly improves feature extraction and OCR accuracy. Option D is correct because low confidence on specific fields usually indicates the model has not seen enough layout variation for those fields, so adding diverse training documents that capture the different layouts, fonts, and positions where those fields appear lets the model generalize and raise its confidence. Option A is wrong because the confidence threshold in the analyzeDocument request only filters which results are returned or flagged; it does not change the model's actual extraction confidence. Option B is wrong because converting to grayscale and downsampling to 72 DPI reduces image fidelity and typically lowers OCR and extraction quality rather than improving confidence. Option E is wrong because a prebuilt business card model is a fixed, general-purpose model that cannot be trained on your specific fields or layouts, so it cannot improve confidence for your custom template fields.

Answer

Increase the confidence threshold parameter in the analyzeDocument request body to 0.99.

Answer

Convert all training PDFs to grayscale and reduce their resolution to 72 DPI to speed up processing.

Answer

Switch the model type from custom template to a prebuilt business card model.

Question 8

You are configuring an Azure AI Document Intelligence custom extraction pipeline that processes thousands of diverse medical forms daily. You notice that extraction confidence scores for certain hand-written fields are low. What is the most effective remediation step within the Document Intelligence workflow?

Accepted Answer

Add additional labeled training documents that specifically feature diverse handwriting samples for those low-confidence fields.. The most effective remediation for low-confidence extraction of handwritten fields in a custom Document Intelligence model is to provide more labeled training data that represents the variability causing the low confidence. Handwriting is highly variable, so adding diverse handwriting samples for the specific fields helps the model learn the patterns and improve accuracy. This directly addresses the root cause—insufficient training examples for those field variations—within the Document Intelligence workflow.

Answer

Disable the OCR engine parameter in the API call to bypass text recognition steps.

Answer

Decrease the model training iterations configuration setting to prevent overfitting on handwriting.

Answer

Switch from a Custom Neural model to a Custom Template model.

Question 9

You are designing a multi-tenant SaaS application on Azure. Each tenant requires isolation of their Azure OpenAI token usage and cost tracking. What is the recommended provisioning strategy to achieve strict cost and quota isolation per tenant?

Accepted Answer

Provision a separate Azure OpenAI resource for each tenant.. Provisioning a separate Azure OpenAI resource per tenant provides independent quota limits, separate keys, and distinct billing meters, which together deliver strict cost and quota isolation. This is the recommended multi-tenant pattern when isolation is a hard requirement.

Answer

Use a single Azure OpenAI resource for all tenants and track usage using custom HTTP headers.

Answer

Deploy all tenants in the F0 free pricing tier.

Answer

Rely on Azure Cost Management resource group tags alone without resource separation.

Question 10

You are building a custom named entity recognition model in Language Studio. You want to extract product serial numbers that follow a specific pattern (two uppercase letters followed by four digits, a hyphen, and three digits). Which entity type definition method should you use?

Accepted Answer

Regex entity type. A regex entity type is designed to extract entities that match a specific regular expression pattern, such as two uppercase letters followed by four digits, a hyphen, and three digits (e.g., '[A-Z]{2}\d{4}-\d{3}'). In Language Studio, you define the regex pattern and optionally provide examples, making it ideal for structured identifiers like serial numbers.

Answer

List entity type

Answer

Learned entity type

Answer

Prebuilt entity type

Question 11

You are implementing prompt engineering for a customer support agent. To minimize hallucinations when the model lacks information, which Prompt Flow technique is most effective?

Accepted Answer

Instructional grounding to only answer from context. Instructional grounding — explicitly instructing the model to answer only from the provided context and to say it doesn't know otherwise — is the most direct and effective prompt-engineering technique to minimize hallucinations when the model lacks information. It constrains the model's generation to the retrieved context rather than its parametric memory.

Answer

Chain-of-thought prompting

Answer

Increasing the Top_P parameter

Answer

Few-shot prompting with varied examples

Question 12

You are designing an agentic workflow in Azure AI Foundry that requires the agent to call an external REST API. Which component should you use to define the tool interface?

Accepted Answer

Tool definition using a Python function. In Azure AI Foundry agentic workflows, tools are defined as callable functions with a name, description, and parameter schema; a Python function definition (often decorated or registered as a tool) is the standard way to expose an external REST API to the agent. The model then emits structured arguments that invoke that function. This is the correct component for defining the tool interface.

Answer

Azure AI Content Safety

Answer

LLM output parser

Answer

Flow input schema

Question 13

You are developing an application that processes multilingual documents using Azure AI Language. You want to ensure that your API payload complies with the maximum document size limit per single request document in synchronous calls. What is the maximum character length allowed for a single document input?

Accepted Answer

5,120 characters. Azure AI Language synchronous API calls accept a maximum of 5,120 characters per document in a single request. This limit applies to synchronous document analysis operations and must be respected to avoid request failures.

Answer

1,000,000 characters

Answer

50,000 characters

Answer

1,024 characters

Question 14

You are troubleshooting an Azure AI Language asynchronous summarization job. You submitted a batch of documents using the `/analyze` endpoint with long-running operations (LRO). How should your application check the status and retrieve the results of the job?

Accepted Answer

Poll the URL provided in the 'operation-location' response header using a GET request until the status is 'succeeded'.. Azure AI Language asynchronous (LRO) jobs return an 'operation-location' header in the initial 202 Accepted response. The application must poll that URL with GET requests until the status field reads 'succeeded', then retrieve results from the same endpoint. This is the standard Azure long-running operation pattern.

Answer

Listen for incoming webhook notifications on an Azure Event Grid subscription attached to the text analytics resource.

Answer

Query the Azure Monitor metrics API for job completion status.

Answer

Repeatedly send synchronous POST requests to the same document endpoint until a 200 OK is returned.

Question 15

You have an Azure AI application that is hitting rate limits (429 errors). You need to determine which specific client application is causing the issue. How do you identify this?

Accepted Answer

Enable diagnostic logging to Log Analytics and run a KQL query.. Diagnostic logging to Log Analytics captures per-request telemetry including the caller identity, subscription, resource, operation name, and HTTP status code, so a KQL query filtering on ResponseCode == 429 and summarizing by CallerIpAddress or identity reveals which client is exhausting the quota. The Activity Log only records control-plane operations (resource create/update/delete), not data-plane inference calls, so it cannot show which app is making the requests.

Answer

Check the 'Activity Log'.

Answer

Increase the resource quota in the 'Quotas' blade.

Answer

Restart the AI service.

Question 16

Which THREE of the following are components of a standard Prompt Flow DAG (Directed Acyclic Graph)?

Accepted Answer

Inputs. In a standard Prompt Flow DAG, Inputs (A) are the declared entry points that supply values (for example, a question string or a dataset column) to the flow, making them a required component. Outputs (B) are also required because they define the flow's returned values (such as the model's answer or a score) that downstream consumers or evaluations read. Tools (D) are the executable nodes of the DAG — LLM, Python, Prompt, and built-in tools — and the edges connecting them define the directed acyclic execution order. Model Training Jobs (C) are not part of a Prompt Flow DAG; training is handled by separate Azure Machine Learning job constructs, not flow nodes. Data Labeling Projects (E) are likewise a distinct Azure Machine Learning feature for annotating data and are not components of a Prompt Flow DAG.

Answer

Model Training Jobs

Answer

Data Labeling Projects

Question 17

You are configuring a secure enterprise architecture where Azure AI Vision calls must not traverse the public internet. You create a private endpoint for your Azure AI Services resource in your Virtual Network. Which DNS configuration step is required for client applications within the VNet to resolve the private endpoint correctly?

Accepted Answer

Configure Azure Private DNS zones corresponding to the cognitive services DNS zone and link it to your VNet.. When you create a private endpoint for an Azure AI Services resource, the private IP is associated with the service's public DNS name (for example, <resource>.cognitiveservices.azure.com). For clients in the VNet to resolve that name to the private IP instead of the public IP, you must create an Azure Private DNS zone named privatelink.cognitiveservices.azure.com and link it to the VNet, with an A record mapping the resource to the private endpoint IP.

Answer

Add the endpoint IP to the local hosts file of every Azure AI Vision instance.

Answer

Configure public DNS records on your corporate DNS server to point to the public IP.

Answer

Modify the routing table to drop all DNS traffic.

Question 18

Which THREE attributes are critical for a successful custom model training set?

Accepted Answer

High-quality images. Option A (High-quality images) is correct because custom model training relies on OCR and layout analysis, so clear, well-lit, high-resolution scans or photos with minimal skew, blur, and noise give the model reliable visual features to learn from. Option B (Consistent labeling across all documents) is correct because supervised training requires ground-truth labels to be applied with the same schema and conventions throughout the dataset; inconsistent labels introduce contradictory signals that degrade accuracy and confuse the model. Option C (Representative samples of all document layouts) is correct because the model must generalize to every form, template, and layout it will encounter in production, so the training set must include examples covering the full variety of structures, fonts, and field placements. Option D is not required because documents need not come from a single printer; in fact, variety in printers, scanners, and capture conditions improves robustness. Option E is not required because there is no fixed minimum of 1,000 pages per model; the needed volume depends on document complexity and diversity, and quality and representativeness matter more than a rigid page count.

Answer

All documents from the same printer

Answer

At least 1,000 pages per model

Question 19

You are designing an enterprise RAG solution using Azure AI Search and Azure OpenAI. You need to ensure data security and compliance across different user permission tiers. Which TWO actions should you implement? (Choose two.)

Accepted Answer

Ensure all data transfers between Azure AI Search and Azure OpenAI utilize TLS 1.2 or higher.. Option A is correct because Azure AI Search and Azure OpenAI communicate over HTTPS endpoints, and enforcing TLS 1.2 or higher protects data in transit against interception or downgrade attacks, which is a baseline requirement for enterprise security and compliance. Option B is correct because Azure AI Search supports document-level access control via security filters (for example, using the user's group IDs in a filterable field or the built-in ACL/rbacScope fields), ensuring each permission tier only retrieves documents its users are authorized to see, which is essential in a multi-tenant RAG solution. Option C is wrong because storing unencrypted authentication tokens in embedding metadata exposes credentials to anyone querying the index and violates token-handling best practices. Option D is wrong because granting admin privileges to all users removes least-privilege enforcement and does not address retrieval latency in a legitimate way. Option E is wrong because disabling managed identities and embedding plaintext connection strings in source code leaks secrets and discards Azure's keyless, identity-based authentication model.

Answer

Store unencrypted user authentication tokens directly inside the embedding vector metadata.

Answer

Grant every user application admin privileges to bypass retrieval latency checks.

Answer

Disable managed identities and store plaintext connection strings in the application source code.

Question 20

You are implementing opinion mining as part of sentiment analysis in Azure AI Language. An incoming review states, 'The room was spacious, but the staff was rude.' What does opinion mining return for this sentence?

Accepted Answer

Aspect targets ('room', 'staff') along with their corresponding evaluations ('spacious', 'rude') and individual sentiments.. Opinion mining in Azure AI Language returns aspect-based sentiment analysis, identifying aspect targets (e.g., 'room', 'staff') and their corresponding evaluations (e.g., 'spacious', 'rude') with individual sentiment polarities. For the given review, it would return 'room' as positive and 'staff' as negative, providing granular insights beyond document-level sentiment.

Answer

A binary classification indicating whether the review is spam or genuine.

Answer

A single neutral sentiment score for the entire document.

Answer

A translated version of the review in English.