AB-620 overview

  • 200 practice questions
  • 33 flashcards
  • 5 study guide topics
  • 40 key terms

What you'll study

  • Plan and design AI solutions20%
  • Build and configure AI agents and workflows25%
  • Integrate AI solutions with data and services20%
  • Test, monitor, and optimize AI solutions20%
  • Secure, govern, and deploy AI solutions15%

AB-620 is a professional certification exam focused on building and operating AI solutions on a modern cloud platform. It suits developers, solution architects, and technical consultants who design agents, connect them to real data and services, and then keep them reliable in production. The exam is organized around five domains: planning and designing AI solutions, building and configuring agents and workflows, integrating with data and services, testing and monitoring, and securing, governing, and deploying. Expect scenario-driven questions rather than pure recall: you are given a business constraint or a failing behavior and asked to choose the most appropriate design or remediation. Strong preparation means hands-on practice-building an agent end to end, wiring it to a data source and an API, adding evaluation and logging, and applying access controls and deployment safeguards. Supplement that with the vendor's official documentation and learning paths for the current exam objectives. This page is independent study material and is not affiliated with or endorsed by the certifying body; always confirm the current skills outline and registration details on the official site before you schedule.

Common mistakes

Designing an agent solution without first writing down the business outcome, success criteria, and constraints, then jumping straight to choosing a model or agent topology.

Candidates treat design questions as product-selection questions, so they pick a fashionable pattern (multi-agent orchestration, autonomous tool use) before establishing what the workflow must achieve and what limits apply.

What to do instead

For every design scenario, first state the goal, measurable success criteria, data sensitivity, latency/cost budget, and human-oversight needs; only then select a pattern (single agent, deterministic workflow, RAG, or classic automation) and justify it against those constraints.

Choosing a multi-agent or fully autonomous design for a task that a single agent or a fixed deterministic workflow would satisfy more reliably.

Multi-agent designs sound more capable, and candidates conflate complexity with quality, ignoring added failure surface, coordination cost, and debugging difficulty.

What to do instead

Apply the simplest design that meets the requirements; reserve multi-agent decomposition for genuinely separable subtasks with clear handoffs, and prefer deterministic orchestration when the steps are known in advance.

Defining agent instructions and tool schemas vaguely (e.g., 'search the knowledge base' or untyped parameters), then expecting the agent to behave consistently.

Candidates focus on the happy-path narrative and underestimate how much model behaviour depends on precise instructions, explicit tool descriptions, parameter schemas, and stated boundaries.

What to do instead

Write instructions that specify role, scope, step order, refusal conditions, and output format; give each tool a descriptive name, typed parameters with enums where possible, and clear guidance on when and when not to call it.

Building an agent loop with no termination condition, iteration cap, or fallback path when a tool call fails or returns nothing useful.

Candidates validate only successful runs and do not reason about repeated tool calls, empty results, or ambiguous model output, which produce loops or silent failures.

What to do instead

Design explicit exit criteria, a maximum step or retry count, error handling per tool, and a defined escalation or graceful-degradation path; describe these in any workflow design answer.

Treating retrieval as a plug-in step without designing chunking, metadata, refresh, and relevance thresholds, then blaming the model when answers are wrong.

Candidates assume 'add a vector store' equals grounded answers and skip decisions about document granularity, filtering, and how stale or conflicting content is handled.

What to do instead

Design retrieval deliberately: choose chunk size and overlap for the content type, attach metadata for filtering and citation, define indexing/refresh cadence, set similarity or reranking thresholds, and specify behaviour when no relevant content is found.

Connecting an agent to backend systems with broad credentials and no scoping, validation, or confirmation step for write operations.

Candidates concentrate on making the integration work and treat authorization, input validation, and irreversible actions as afterthoughts.

What to do instead

Give each integration least-privilege access, validate and sanitize parameters before execution, require human confirmation for destructive or high-impact actions, and log every call with the identity and inputs used.

Evaluating an AI solution only with a handful of manual happy-path prompts and declaring it ready.

Candidates rely on informal spot checks because building a representative evaluation set takes effort, so edge cases, regressions, and quality drift go undetected.

What to do instead

Define an evaluation dataset covering typical, edge, adversarial, and refusal cases with expected outcomes; measure groundedness, task success, and safety metrics; rerun the suite on every prompt, model, or data change.

Monitoring only availability and latency while ignoring output quality, token cost, tool-failure rates, and user feedback signals.

Candidates map monitoring to traditional uptime metrics and overlook that an AI system can be 'up' while producing degraded or unsafe answers.

What to do instead

Instrument the full pipeline: log prompts, retrieved context, tool calls, token usage, latency, error and refusal rates, and user feedback; define quality and cost thresholds with alerts and a review loop for sampled outputs.

Handling security only at the network or identity layer while leaving prompt injection, data leakage, and unsafe tool use unaddressed.

Candidates apply conventional app-security thinking and do not treat model input, retrieved content, and tool outputs as untrusted data that can influence agent behaviour.

What to do instead

Treat all external content as untrusted, separate instructions from data, constrain tool permissions, filter and validate model outputs, apply content-safety controls, and test with injection and exfiltration scenarios.

Describing deployment as 'publish the agent' without specifying environment promotion, versioning, rollback, and governance controls such as logging, access review, and data-handling policy.

Candidates conflate a working prototype with a production release and omit the operational and compliance scaffolding the exam expects.

What to do instead

Define separate dev/test/prod environments with promotion gates, version prompts, models, and configurations, plan rollback, and specify audit logging, access control, retention, and responsible-AI review as part of the deployment design.

Exam requirements can change. Confirm current details with the exam provider.

AB-620 practice

200 questions. Each one lists the choices, the correct answer, and an explanation. With JavaScript, this list opens one question at a time. The checker at the top of this page saves progress in this browser when it is available.

  1. Question 1Plan and design AI solutions

    An organization wants a solution that automatically extracts structured fields from scanned invoices. Which Azure AI service is purpose-built for this scenario?

    1. Azure AI Language
    2. Azure AI Vision
    3. Azure AI Speech
    4. Azure AI Document Intelligence Correct answer.

    Explanation

    Document Intelligence provides prebuilt and custom models that extract fields, tables, and key-value pairs from documents such as invoices and receipts. Vision handles image analysis, Language handles text analytics, and Speech handles audio.

  2. Question 2Plan and design AI solutions

    A team needs a retrieval-augmented generation pattern where the model answers questions using their own documents. Which design best supports this requirement?

    1. Store documents in a search index and ground responses with retrieved passages Correct answer.
    2. Fine-tune a base model on the full document set for every update
    3. Enable content filtering to inject document text into prompts
    4. Increase the model temperature so it can recall more content

    Explanation

    RAG grounds model output by retrieving relevant passages from an indexed knowledge source and adding them to the prompt. This keeps answers current without retraining and reduces hallucination.

  3. Question 3Plan and design AI solutions

    Which Azure OpenAI capability allows a model to call external functions or APIs based on user intent?

    1. Provisioned throughput
    2. Function calling Correct answer.
    3. Embeddings
    4. Content filtering

    Explanation

    Function calling lets the model emit structured arguments that your application uses to invoke external tools or APIs. Embeddings produce vectors, filtering screens content, and provisioned throughput is a capacity model.

  4. Question 4Plan and design AI solutions

    A solution must keep all prompt and response data within a specific geographic region for compliance. Which design decision most directly satisfies this requirement?

    1. Deploy resources in a region that meets the residency requirement Correct answer.
    2. Enable diagnostic logging to a global workspace
    3. Set the model temperature to zero
    4. Use a larger model tier

    Explanation

    Data residency is controlled by the region in which the Azure resources are deployed. Model parameters and tiers do not affect where data is processed, and global logging can conflict with residency goals.

  5. Question 5Plan and design AI solutions

    When designing a solution that uses Azure OpenAI, which factor most influences whether to use the pay-as-you-go model versus provisioned throughput?

    1. The number of prompt templates stored
    2. The programming language used by the client
    3. The number of developers on the team
    4. The predictability and volume of token consumption Correct answer.

    Explanation

    Provisioned throughput is cost-effective for steady, high-volume workloads, while pay-as-you-go suits variable or low usage. Consumption patterns are the primary driver of this choice.

  6. Question 6Plan and design AI solutions

    Which Azure AI service should be used to build a conversational agent that answers FAQ questions from a knowledge base with minimal custom code?

    1. Azure AI Search with semantic ranking only
    2. Azure AI Bot Service with a language understanding model Correct answer.
    3. Azure AI Translator
    4. Azure AI Content Safety

    Explanation

    Azure AI Bot Service combined with language understanding supports building conversational agents that resolve intents and answer from knowledge sources. The other services provide search, moderation, or translation rather than conversational orchestration.

  7. Question 7Plan and design AI solutions

    A design requires that user prompts never be used to improve the underlying foundation model. Which approach addresses this concern?

    1. Disable content filtering
    2. Rely on the default model behavior without configuration
    3. Use an Azure OpenAI deployment where data is not used for model training Correct answer.
    4. Store prompts in a public storage account

    Explanation

    Azure OpenAI deployments do not use customer prompts or completions to train or improve foundation models. Public storage and disabling filtering do not address training data concerns.

  8. Question 8Plan and design AI solutions

    Which design pattern best reduces hallucination when an AI solution must answer questions about a regulated policy corpus?

    1. Use the smallest available model to reduce cost
    2. Raise the temperature to encourage varied answers
    3. Shorten the system prompt to save tokens
    4. Ground responses in retrieved source passages and require citations Correct answer.

    Explanation

    Grounding answers in retrieved passages and requiring citations ties output to verifiable sources and reduces fabrication. Higher temperature and smaller models can increase inaccuracy, and prompt length is not the cause of hallucination.

  9. Question 9Plan and design AI solutions

    A solution must classify incoming support tickets into categories and route them. Which combination of Azure AI services fits best?

    1. Azure AI Content Safety and Azure AI Search
    2. Azure AI Language custom text classification and an orchestration layer Correct answer.
    3. Azure AI Speech and Azure AI Translator
    4. Azure AI Vision and Azure AI Document Intelligence

    Explanation

    Custom text classification in Azure AI Language assigns categories to text, and an orchestration layer routes based on the result. The other combinations address unrelated modalities or functions.

  10. Question 10Plan and design AI solutions

    Which statement best describes the role of embeddings in a semantic search design?

    1. They enforce rate limits on API calls
    2. They represent text as vectors so similarity can be computed Correct answer.
    3. They translate queries into multiple languages
    4. They compress documents to reduce storage cost

    Explanation

    Embeddings map text into a vector space where semantically similar content is close together, enabling similarity search. They are not a compression, translation, or throttling mechanism.

  11. Question 11Plan and design AI solutions

    A team is designing a multi-stage pipeline that ingests documents, extracts text, chunks it, generates embeddings, and indexes it. Where should chunking occur?

    1. After indexing in the search service
    2. Only at query time
    3. Before text extraction
    4. After text extraction and before embedding generation Correct answer.

    Explanation

    Chunking happens after text is extracted and before embeddings are generated so each chunk can be embedded and indexed. Indexing and query time occur downstream of embedding.

  12. Question 12Plan and design AI solutions

    Which Azure feature helps protect an AI application from prompt injection attacks?

    1. Azure Monitor metrics
    2. Azure AI Translator
    3. Prompt Shields in Azure AI Content Safety Correct answer.
    4. Azure Key Vault

    Explanation

    Prompt Shields detect and block attempts to manipulate a model through injected instructions. Translator, metrics, and Key Vault serve unrelated purposes.

  13. Question 13Plan and design AI solutions

    A design requires that only authorized employees can query an internal AI assistant, and access must be auditable. Which approach best meets this?

    1. Disable logging to reduce storage costs
    2. Authenticate users with Microsoft Entra ID and log access Correct answer.
    3. Rely on network isolation alone
    4. Embed a shared API key in the client application

    Explanation

    Identity-based authentication with Entra ID enforces per-user authorization, and logging provides an audit trail. Shared keys and network isolation alone do not identify individual users, and disabling logging removes auditability.

  14. Question 14Plan and design AI solutions

    Which scenario is the strongest candidate for fine-tuning rather than retrieval augmentation?

    1. Answering questions over a knowledge base that changes weekly
    2. Enforcing data residency requirements
    3. Teaching a consistent output style and format for a fixed task Correct answer.
    4. Looking up the current status of a support ticket

    Explanation

    Fine-tuning adapts model behavior, tone, and format for a stable task. Rapidly changing knowledge is better served by retrieval, live lookups require tool calls, and residency is a deployment concern.

  15. Question 15Plan and design AI solutions

    An AI solution must process both text and images in a single request. Which model capability should be selected?

    1. A speech-to-text model
    2. A text-only completion model
    3. A multimodal model Correct answer.
    4. An embeddings-only model

    Explanation

    Multimodal models accept more than one input modality, such as text and images, in a single request. Text-only, embeddings-only, and speech models each handle a narrower input type.

  16. Question 16Plan and design AI solutions

    A design must limit the blast radius of a compromised API key used by an AI application. Which practice best achieves this?

    1. Share the key across all environments
    2. Store the key in source code for reliability
    3. Use managed identity and rotate secrets stored in Key Vault Correct answer.
    4. Give the key full subscription-level permissions

    Explanation

    Managed identities remove embedded secrets, and rotating secrets in Key Vault limits exposure if a key is compromised. Hardcoding, over-privileging, and sharing keys widen the blast radius.

  17. Question 17Plan and design AI solutions

    Which design consideration is most important when choosing a chunk size for a retrieval index?

    1. Balancing context completeness against retrieval precision Correct answer.
    2. Maximizing the number of chunks regardless of content
    3. Ensuring chunks are always a single word
    4. Setting chunk size equal to the model's maximum token limit

    Explanation

    Chunk size trades off completeness of context against precision of retrieval; chunks that are too large dilute relevance and too small lose context. Maximizing count or using extreme sizes harms quality.

  18. Question 18Plan and design AI solutions

    A solution must moderate user-generated text for hate, violence, and self-harm before display. Which service should be integrated?

    1. Azure AI Content Safety Correct answer.
    2. Azure AI Speech
    3. Azure AI Translator
    4. Azure AI Document Intelligence

    Explanation

    Azure AI Content Safety classifies harmful text and imagery across categories such as hate, violence, and self-harm. The other services handle documents, speech, and translation.

  19. Question 19Plan and design AI solutions

    An organization wants to deploy an AI assistant that must remain functional if a single Azure region experiences an outage. Which design element addresses this?

    1. Deploy resources across multiple regions with failover Correct answer.
    2. Increase the model temperature
    3. Reduce the token limit per request
    4. Deploy all resources to one region with a single replica

    Explanation

    Multi-region deployment with failover provides resilience against regional outages. Single-region, single-replica designs have a single point of failure, and model parameters do not affect availability.

  20. Question 20Plan and design AI solutions

    Which metric is most useful for evaluating whether a retrieval-augmented solution returns relevant context before generation?

    1. Model temperature setting
    2. Token count of the final answer
    3. Number of API deployments
    4. Recall of relevant passages in the retrieved set Correct answer.

    Explanation

    Retrieval quality is measured by how many relevant passages appear in the retrieved set, commonly recall or precision at k. Answer length, temperature, and deployment count do not measure retrieval quality.

  21. Question 21Plan and design AI solutions

    A design requires processing large volumes of documents asynchronously without blocking user requests. Which approach fits best?

    1. Queue-based background processing with a document pipeline Correct answer.
    2. Synchronous calls inside the user request
    3. Disabling retries to speed up processing
    4. Manual processing by administrators

    Explanation

    Queue-based background processing decouples long-running document work from user requests and supports scaling. Synchronous in-request processing increases latency, and removing retries reduces reliability.

  22. Question 22Plan and design AI solutions

    Which statement best explains why a system prompt should be designed to resist instruction override?

    1. System prompts reduce token costs
    2. System prompts replace the need for content filtering
    3. System prompts establish behavior that user input should not be able to replace Correct answer.
    4. System prompts are visible to all users by default

    Explanation

    System prompts set the model's role and constraints, and robust design helps prevent user input from overriding those constraints. They are not a cost tool or a substitute for content filtering.

  23. Question 23Plan and design AI solutions

    A team wants to compare model outputs against a human-labeled dataset to measure answer quality. Which practice should they adopt?

    1. Increase temperature until outputs look correct
    2. Define evaluation metrics and run a repeatable test set Correct answer.
    3. Skip evaluation if the model is popular
    4. Rely on a single anecdotal example

    Explanation

    Repeatable evaluation against a labeled test set with defined metrics gives objective quality measures. Anecdotes, tuning until outputs look right, and popularity are not valid evaluation methods.

  24. Question 24Plan and design AI solutions

    An AI solution must ensure that responses cite only documents the requesting user is permitted to see. Which design element is essential?

    1. Caching all documents in a shared index without filters
    2. Applying user-specific access filters during retrieval Correct answer.
    3. Using a larger embedding dimension
    4. Increasing the model's context window

    Explanation

    Access control must be enforced at retrieval time with user-specific filters so the model only receives authorized content. Context size, unfiltered shared indexes, and embedding dimensions do not enforce permissions.

  25. Question 25Plan and design AI solutions

    Which Azure service is best suited to convert recorded call audio into searchable text for later analysis?

    1. Azure AI Document Intelligence
    2. Azure AI Speech Correct answer.
    3. Azure AI Vision
    4. Azure AI Content Safety

    Explanation

    Azure AI Speech provides speech-to-text transcription for audio, enabling downstream search and analysis. The other services process images, moderate content, or extract document fields.

  26. Question 26Plan and design AI solutions

    A company wants to build a solution that summarizes long contracts and must decide between a single large prompt and a map-reduce approach. Which factor most directly justifies choosing the map-reduce pattern?

    1. The documents exceed the model's maximum context window. Correct answer.
    2. The documents must be stored in a relational database.
    3. The documents contain scanned images.
    4. The summaries must be translated into several languages.

    Explanation

    Map-reduce splits a document into chunks that are summarized separately and then combined, which is necessary when the input cannot fit within the model's context window. The other factors relate to OCR, localization, or storage and do not determine the summarization pattern.

  27. Question 27Plan and design AI solutions

    A design must ensure that an AI assistant's responses are grounded in the organization's internal product documentation rather than general model knowledge. Which architectural component is essential?

    1. A content safety filter applied to outputs
    2. A larger model with more parameters
    3. A retrieval step that supplies relevant document excerpts to the model Correct answer.
    4. A caching layer in front of the model endpoint

    Explanation

    Grounding requires that relevant source material be retrieved and included in the model input so answers are based on that material. Model size, output filtering, and caching do not connect responses to the internal corpus.

  28. Question 28Plan and design AI solutions

    A team is planning an AI solution and wants to estimate the cost of model usage before launch. Which quantity should they forecast to predict token-based charges?

    1. The number of user accounts in the directory
    2. The number of virtual machines in the subnet
    3. The number of input and output tokens per request multiplied by expected request volume Correct answer.
    4. The number of storage accounts provisioned

    Explanation

    Token-based pricing depends on how many prompt and completion tokens are processed, so forecasting both per request and multiplying by volume estimates cost. Storage, identities, and compute hosts are unrelated to token consumption.

  29. Question 29Plan and design AI solutions

    A solution must let users ask questions over a knowledge base and receive answers that include links to the source passages. Which design element directly enables the citations?

    1. Storing chunk identifiers and source metadata alongside each embedded chunk Correct answer.
    2. Enabling streaming of the response to the client
    3. Increasing the model temperature setting
    4. Compressing the index with a smaller embedding dimension

    Explanation

    Citations require that each retrieved chunk retain metadata identifying its source so the response can reference it. Temperature, embedding dimension, and streaming affect generation behavior or performance but not traceability to sources.

  30. Question 30Plan and design AI solutions

    An organization needs an AI solution that automatically detects the language of incoming messages and translates them into a single working language before analysis. Which design choice is most appropriate?

    1. Restrict input to a single language at the client
    2. Use a language detection step, then a translation service, before the analysis stage Correct answer.
    3. Store all messages untranslated and translate only when a user asks
    4. Rely on the analysis model to interpret every language without preprocessing

    Explanation

    Explicit detection followed by translation normalizes input so downstream analysis operates on one language consistently. Skipping preprocessing, deferring translation, or restricting input does not meet the requirement to handle multiple languages automatically.

  31. Question 31Plan and design AI solutions

    A design review asks why an AI solution should log the model version, prompt template version, and retrieval configuration for each request. Which goal does this logging primarily serve?

    1. Load balancing across regions
    2. Reproducibility and root-cause analysis of response changes over time Correct answer.
    3. Encryption of data at rest
    4. Reduction of token costs per request

    Explanation

    Recording the model, prompt, and retrieval settings lets teams reproduce and diagnose why outputs changed after a configuration update. These logs do not reduce token usage, encrypt data, or distribute traffic.

  32. Question 32Plan and design AI solutions

    An AI solution must enforce that responses never contain personal data such as account numbers, even if such data appears in retrieved documents. Which design approach best meets this requirement?

    1. Apply a redaction step to retrieved content and to generated output before display Correct answer.
    2. Lower the number of retrieved chunks to one
    3. Store the index in a different region
    4. Increase the model's context window

    Explanation

    Redacting sensitive patterns in retrieved content and again on output provides defense in depth so personal data does not reach the user. Context size, region, and chunk count do not address redaction.

  33. Question 33Plan and design AI solutions

    A team is deciding whether to use a prebuilt model or train a custom model for image classification of product defects. Which consideration is most relevant to the decision?

    1. Whether the defect categories are specific to the business and not covered by general models Correct answer.
    2. Whether the images will be stored in blob storage
    3. Whether the development team prefers Python or C#
    4. Whether the solution will be billed monthly or annually

    Explanation

    Custom training is justified when categories are domain-specific and not reliably recognized by general-purpose models. Billing cadence, programming language, and storage choice do not determine whether a prebuilt model is sufficient.

  34. Question 34Plan and design AI solutions

    A solution must route high-confidence automated decisions directly and send low-confidence cases to a human reviewer. Which design element is required to implement this?

    1. A single shared API key for all reviewers
    2. A confidence threshold evaluated per response with a fallback path to human review Correct answer.
    3. A nightly batch job that reprocesses all records
    4. A fixed temperature of zero for all requests

    Explanation

    A threshold on the model's confidence score determines whether a case is auto-approved or escalated. Temperature, key management, and batch reprocessing do not create a decision boundary for human routing.

  35. Question 35Plan and design AI solutions

    A company wants an AI assistant to answer questions using data that must remain within its own subscription and must not be shared with the model provider for training. Which design consideration is most important?

    1. Enabling verbose logging of all prompts
    2. Choosing an offering with contractual data-use terms that exclude training on customer data Correct answer.
    3. Deploying the assistant in the most inexpensive region
    4. Selecting the model with the largest context window

    Explanation

    Data-handling commitments are established through the service's contractual terms and configuration, not by model size or region. Verbose prompt logging would increase exposure rather than protect it.

  36. Question 36Plan and design AI solutions

    A solution must support a conversational experience where the assistant remembers details from earlier in the same session. Which design element provides this capability?

    1. Increasing the maximum output token limit
    2. Disabling content safety checks for the session
    3. Maintaining a session-scoped conversation history included in each model request Correct answer.
    4. Storing all conversations permanently in the retrieval index

    Explanation

    Short-term memory is achieved by including prior turns of the current session in the request context. Permanent indexing, output limits, and disabling safety do not provide conversational continuity.

  37. Question 37Plan and design AI solutions

    A design specifies that the AI solution must degrade gracefully when the primary model endpoint is throttled. Which approach best satisfies this requirement?

    1. Cache only the system prompt locally
    2. Send requests to the endpoint more frequently to warm it up
    3. Implement retry with backoff and fall back to a secondary model deployment Correct answer.
    4. Remove all rate limiting from the client

    Explanation

    Retry with backoff handles transient throttling, and a secondary deployment provides continuity if the primary remains unavailable. Removing limits, increasing request rate, or caching only the prompt does not prevent failure under throttling.

  38. Question 38Plan and design AI solutions

    A team wants to evaluate whether a planned AI solution will meet latency targets before full build-out. Which approach is most appropriate?

    1. Estimate latency from the model's parameter count alone
    2. Build a prototype that measures end-to-end latency with representative inputs Correct answer.
    3. Defer latency testing until after production deployment
    4. Assume latency equals the sum of documented service limits

    Explanation

    End-to-end latency depends on retrieval, orchestration, and model response time, so a prototype with representative inputs gives reliable measurements. Parameter count, summed limits, and post-deployment testing do not validate the target beforehand.

  39. Question 39Plan and design AI solutions

    A solution must extract structured fields such as invoice number, date, and total from scanned invoices. Which capability should the design rely on?

    1. Text translation of the invoice into another language
    2. Document intelligence that returns key-value pairs from forms Correct answer.
    3. Speech-to-text transcription of the invoice
    4. Sentiment analysis on the invoice text

    Explanation

    Form-oriented document processing is designed to return structured fields from scanned documents. Sentiment, transcription, and translation address different tasks and do not extract invoice fields.

  40. Question 40Plan and design AI solutions

    A design must ensure that a retrieval index stays current as source documents are added, updated, and deleted. Which approach best addresses this?

    1. Manual reindexing requested by users when answers seem stale
    2. A full reindex performed once at initial deployment
    3. An incremental indexing pipeline triggered by document change events Correct answer.
    4. Caching retrieval results indefinitely to reduce load

    Explanation

    Event-driven incremental indexing keeps the index synchronized with source changes, including deletions. One-time reindexing, manual requests, and indefinite caching allow the index to drift from the source.

  41. Question 41Build and configure AI agents and workflows

    In a Microsoft Copilot Studio topic, which element determines the specific user phrases that trigger the topic?

    1. Question nodes
    2. Message nodes
    3. Condition branches
    4. Trigger phrases Correct answer.

    Explanation

    A topic is entered when the user's message matches one of its trigger phrases, which the author supplies as example utterances. Other nodes control conversation flow after the topic is entered, not entry itself.

  42. Question 42Build and configure AI agents and workflows

    You need a value captured in one topic to remain available later in the same conversation so another topic can use it. What should you do with the variable?

    1. Store it in a message node
    2. Reset it at the end of the topic
    3. Convert it to a topic-level variable
    4. Mark it as a global variable Correct answer.

    Explanation

    Global variables persist across topics for the duration of the conversation, so any later topic can read them. Topic-level variables are scoped to a single topic and are not reliable across topic boundaries.

  43. Question 43Build and configure AI agents and workflows

    What is the primary purpose of a knowledge source in an agent?

    1. To define the visual theme of the chat canvas
    2. To supply grounding content the agent can search when answering Correct answer.
    3. To authenticate users before the conversation starts
    4. To schedule when the agent is published

    Explanation

    Knowledge sources provide the content an agent retrieves and cites so answers are grounded in authoritative material. Presentation, publishing schedules, and authentication are handled by other configuration areas.

  44. Question 44Build and configure AI agents and workflows

    An agent should hand a conversation to a human agent when the user explicitly asks for a person. Which configuration supports this?

    1. A shorter conversation timeout value
    2. A higher creativity setting on the model
    3. An additional trigger phrase on the greeting topic
    4. A handoff or escalate action configured for the topic Correct answer.

    Explanation

    Handoff is implemented by an explicit escalation path, such as a topic with a handoff action that transfers the conversation to a live queue. Model creativity, triggers, and timeouts do not route work to people.

  45. Question 45Build and configure AI agents and workflows

    A workflow must run automatically whenever a new item is added to a SharePoint list. Which trigger type fits this requirement?

    1. An automated trigger tied to the list event Correct answer.
    2. A manual trigger started by the user
    3. A scheduled recurrence trigger
    4. A button trigger inside a chat client

    Explanation

    Automated triggers fire in response to a defined event such as item creation, which matches the requirement. Manual, scheduled, and button triggers require a person or a clock rather than the event itself.

  46. Question 46Build and configure AI agents and workflows

    Which action type in a cloud flow should you use to call an agent and use its response in later steps?

    1. A compose action that builds a string
    2. A terminate action that ends the run
    3. A delay action that pauses execution
    4. An agent action that returns output to the flow Correct answer.

    Explanation

    Agent actions invoke an agent and expose its response as dynamic content for subsequent steps. Compose, delay, and terminate actions do not call an agent.

  47. Question 47Build and configure AI agents and workflows

    What does an agent's system message primarily control?

    1. The retention period for conversation transcripts
    2. The agent's persona, tone, and behavioral boundaries Correct answer.
    3. The network region where the service runs
    4. The list of channels where the agent is published

    Explanation

    The system message sets instructions that shape how the agent behaves, including tone, scope, and constraints. Channels, retention, and hosting region are governed by separate settings.

  48. Question 48Build and configure AI agents and workflows

    You want an agent to answer only from approved internal documents and decline unrelated questions. Which setting most directly enforces this boundary?

    1. Increasing the number of trigger phrases on every topic
    2. Restricting the agent to its configured knowledge and instructing it to decline off-topic requests Correct answer.
    3. Adding more channels to the agent
    4. Enabling a longer session timeout

    Explanation

    Grounding to configured knowledge plus explicit instructions to decline out-of-scope requests keeps answers within approved material. Triggers, timeouts, and channels do not constrain the content of answers.

  49. Question 49Build and configure AI agents and workflows

    A topic collects an order number, then must branch to different messages depending on whether the number exists in a data source. Which combination achieves this?

    1. A question node storing the value, followed by a condition node that tests the lookup result Correct answer.
    2. A message node followed by a redirect to another topic
    3. Two trigger phrases on the same topic
    4. An end conversation node placed before the question

    Explanation

    The question node captures the value into a variable, and a condition node evaluates it or a lookup result to choose a branch. Messages, triggers, and early termination do not provide branching logic.

  50. Question 50Build and configure AI agents and workflows

    Users report the agent answers general questions well but never starts the topic that handles refunds. What should you check first?

    1. Whether the agent has a handoff action configured
    2. Whether the refund topic's trigger phrases match how users actually phrase requests Correct answer.
    3. Whether the environment has multiple channels enabled
    4. Whether the refund topic has too many condition nodes

    Explanation

    A topic that never starts usually has trigger phrases that do not match real user wording, so the utterance is routed elsewhere. Conditions only matter after the topic begins, and handoff or channels do not affect topic matching.

  51. Question 51Build and configure AI agents and workflows

    Which practice best improves the reliability of an agent that retrieves answers from a large document library?

    1. Increasing the agent's response length limit
    2. Curating source documents into focused, well-structured content and removing outdated duplicates Correct answer.
    3. Adding every available document regardless of relevance
    4. Disabling citations to simplify responses

    Explanation

    Retrieval quality depends heavily on source quality; focused, current, well-structured content yields more accurate grounding. Volume without curation and hiding citations reduce trust and accuracy.

  52. Question 52Build and configure AI agents and workflows

    A flow must call an agent for each item returned by a list query. Where should the agent action be placed?

    1. Before the trigger as a pre-step
    2. Inside a parallel branch that never joins
    3. Inside an apply-to-each loop over the query results Correct answer.
    4. After a terminate action

    Explanation

    Iterating over the results with an apply-to-each loop lets the agent action run once per item. Placing it before the trigger is not possible, unjoined parallel branches complicate results, and steps after terminate never run.

  53. Question 53Build and configure AI agents and workflows

    What is the effect of configuring an authentication option on a channel for an agent?

    1. The agent bypasses its knowledge sources
    2. The conversation is stored permanently without retention limits
    3. Users must sign in before interacting, and the agent can use their identity Correct answer.
    4. The agent automatically translates messages into all languages

    Explanation

    Channel authentication requires users to sign in, which both restricts access and lets the agent personalize responses using the authenticated identity. It does not change translation, grounding, or retention behavior.

  54. Question 54Build and configure AI agents and workflows

    An agent should confirm a destructive action with the user before performing it. Which design satisfies this requirement?

    1. Log the action to a table after it completes
    2. Insert a confirmation question node and act only on an affirmative response Correct answer.
    3. Rely on the model to decide silently whether to proceed
    4. Perform the action and notify the user afterward

    Explanation

    Explicit confirmation requires capturing the user's choice and branching so the action runs only on agreement. Silent model decisions, after-the-fact notices, and logging do not prevent unintended actions.

  55. Question 55Build and configure AI agents and workflows

    Which configuration allows an agent to be reached from a Teams channel and a custom website simultaneously?

    1. Enabling a single channel with a longer timeout
    2. Publishing the agent to multiple channels Correct answer.
    3. Restricting the agent to authenticated users only
    4. Creating duplicate agents per channel

    Explanation

    A single agent can be published to several channels at once, so the same content serves each surface. Duplicating agents creates maintenance overhead, and timeouts or auth restrictions do not add channels.

  56. Question 56Build and configure AI agents and workflows

    A workflow must pause until a human approves a request before continuing. Which pattern implements this?

    1. A delay action set to a fixed number of minutes
    2. A recurrence trigger that runs hourly
    3. A compose action that formats the request
    4. An approval action that waits for a response, followed by a condition on the outcome Correct answer.

    Explanation

    An approval action suspends the flow until a person responds, and a condition routes on the decision. Fixed delays, formatting steps, and schedules do not capture a human decision.

  57. Question 57Build and configure AI agents and workflows

    Which statement best describes how an agent uses generative answers when a matching topic does not exist?

    1. It always returns a fixed fallback string
    2. It escalates to a human automatically
    3. It can generate a response grounded in configured knowledge sources Correct answer.
    4. It restarts the conversation from the greeting

    Explanation

    When no topic matches, generative answers can draw on configured knowledge to answer the question. A fixed fallback, automatic escalation, or restart are separate behaviors that must be configured explicitly.

  58. Question 58Build and configure AI agents and workflows

    You are told the agent sometimes cites an outdated policy document that was replaced last quarter. What is the most direct remediation?

    1. Remove or update the outdated document in the knowledge source Correct answer.
    2. Increase the model's creativity setting
    3. Add more trigger phrases to the policy topic
    4. Shorten the conversation timeout

    Explanation

    Stale citations trace back to stale content in the knowledge source, so updating or removing the document fixes the root cause. Creativity, triggers, and timeouts do not change which documents are retrieved.

  59. Question 59Build and configure AI agents and workflows

    Which approach best handles a user who provides an invalid email address during a form-like topic?

    1. End the conversation immediately
    2. Accept any value and continue silently
    3. Redirect to the greeting topic on the first attempt
    4. Use entity or validation settings that re-prompt until a valid value is entered Correct answer.

    Explanation

    Validation with re-prompting keeps the conversation moving while ensuring the captured value meets the expected format. Silent acceptance, abrupt termination, and restarting all degrade the experience.

  60. Question 60Build and configure AI agents and workflows

    A flow's agent action intermittently fails when the agent takes too long to respond. Which change most directly addresses this?

    1. Publish the agent to an additional channel
    2. Increase the number of trigger phrases on the agent
    3. Add retry or error-handling logic around the agent action Correct answer.
    4. Remove all conditions from the flow

    Explanation

    Transient latency failures are handled with retry policies and scoped error handling so the flow can recover. Removing conditions, changing triggers, or adding channels does not address timeouts.

  61. Question 61Build and configure AI agents and workflows

    What is the main benefit of testing an agent with the built-in test pane before publishing?

    1. It disables generative answers during testing
    2. It automatically rewrites trigger phrases for you
    3. It permanently publishes the agent to all channels
    4. You can trace topic matching and variable values during a simulated conversation Correct answer.

    Explanation

    The test pane shows which topics are triggered and how variables are populated, which helps diagnose routing and logic issues. It does not auto-rewrite content, publish, or disable generative answers.

  62. Question 62Build and configure AI agents and workflows

    An agent must reference a customer's account balance from a line-of-business API during a conversation. Which configuration pattern fits?

    1. Publish the agent to a new channel
    2. Store the balance in a message node
    3. Add the API URL as a trigger phrase
    4. Define a tool or action that calls the API and returns the balance to the agent Correct answer.

    Explanation

    External data is retrieved by defining a tool or action the agent can invoke, which returns structured results for the response. URLs are not triggers, messages cannot fetch data, and channels do not integrate APIs.

  63. Question 63Build and configure AI agents and workflows

    Two topics have overlapping trigger phrases, and the wrong topic frequently starts. What is the most reliable fix?

    1. Delete all condition nodes from both topics
    2. Add a handoff action to both topics
    3. Increase the creativity setting for both topics
    4. Differentiate the trigger phrases so each topic has distinct, representative utterances Correct answer.

    Explanation

    Ambiguous matching stems from overlapping example utterances, so making each topic's triggers distinct and representative resolves the conflict. Conditions, handoff, and creativity do not influence which topic is selected.

  64. Question 64Build and configure AI agents and workflows

    A workflow must process a batch of records and continue past individual failures while collecting errors. Which design is most appropriate?

    1. An apply-to-each loop with a scoped run-after configuration that captures failures and continues Correct answer.
    2. A parallel branch that is never joined
    3. A single agent action placed before the trigger
    4. A terminate action configured to run on the first error

    Explanation

    Scoped error handling inside a loop lets each item fail independently while the flow continues and records errors. Pre-trigger actions are impossible, terminating on first error stops the batch, and unjoined branches complicate aggregation.

  65. Question 65Build and configure AI agents and workflows

    An agent's answers are accurate but it occasionally reveals internal-only content that is present in a connected knowledge source. What is the most effective control?

    1. Add more trigger phrases to the FAQ topic
    2. Restrict the knowledge source to content appropriate for the audience and add instructions limiting disclosure Correct answer.
    3. Increase the response length limit
    4. Enable additional channels for the agent

    Explanation

    If sensitive content is reachable through a knowledge source, the agent can surface it, so the source scope must be corrected and instructions added to constrain disclosure. Length, triggers, and channels do not prevent leakage.

  66. Question 66Build and configure AI agents and workflows

    An agent's maker wants the conversation to begin with a short greeting and a menu of three common requests before any free-form typing. Which component should the maker configure to present those choices?

    1. A system topic that runs on conversation start Correct answer.
    2. A message node inside a trigger phrase list
    3. A fallback topic set to run after two turns
    4. An adaptive card stored in the knowledge source

    Explanation

    System topics such as the conversation-start topic run automatically at defined points in the conversation and are the right place to greet users and offer choices. Trigger phrases only determine when a custom topic is selected and do not fire on their own at startup.

  67. Question 67Build and configure AI agents and workflows

    A maker needs the agent to collect a user's department name and store it for use in several later messages. Which configuration accomplishes this within a topic?

    1. Enable generative answers for the topic
    2. Add a condition node that compares the input to a list
    3. Create a variable and assign the user's response to it Correct answer.
    4. Add a redirect node to the fallback topic

    Explanation

    Variables persist values across nodes and topics, so assigning a user's response to a variable makes that value available to later messages. Conditions compare values but do not themselves capture and retain input.

  68. Question 68Build and configure AI agents and workflows

    A maker wants an agent to hand off to a live support representative when a user types a specific phrase. Which node type provides this behavior?

    1. A transfer conversation node Correct answer.
    2. An adaptive card node with a submit button
    3. A condition node with a branch for the phrase
    4. A message node with a variable reference

    Explanation

    A transfer conversation node ends automated handling and routes the conversation to a configured engagement hub or human queue. Messages and conditions keep the user inside the automated flow.

  69. Question 69Build and configure AI agents and workflows

    Which element of a topic determines the order in which its nodes execute?

    1. The vertical sequence of connected nodes Correct answer.
    2. The order in which the nodes were created
    3. The priority value on each node
    4. The alphabetical order of node titles

    Explanation

    A topic is a flow, and execution follows the connections between nodes from the trigger downward. Creation order and titles have no effect on execution.

  70. Question 70Build and configure AI agents and workflows

    A maker needs the agent to ask a follow-up question only when a user's earlier answer indicates a problem. Which node should be used?

    1. A message node
    2. An end conversation node
    3. A set variable node
    4. A condition node Correct answer.

    Explanation

    Condition nodes evaluate an expression or variable and branch the conversation accordingly, which is how conditional follow-up questions are built. The other nodes act unconditionally.

  71. Question 71Build and configure AI agents and workflows

    What is the purpose of a fallback topic in an agent?

    1. To log every conversation to a data store
    2. To restrict the agent to authenticated users
    3. To handle user input that no other topic matched Correct answer.
    4. To override the system message at runtime

    Explanation

    The fallback topic is triggered when the agent cannot match the user's input to any topic, letting the maker define a graceful response. It does not perform logging, message overrides, or authentication.

  72. Question 72Build and configure AI agents and workflows

    A maker wants to reuse a set of questions about shipping across several topics without rebuilding them each time. Which feature supports this?

    1. Creating a child topic and redirecting to it Correct answer.
    2. Enabling generative answers for each topic
    3. Adding the questions to the system message
    4. Duplicating each topic and editing the copies

    Explanation

    Child topics can be redirected to from multiple parent topics, so shared question sets are authored once and reused. Duplication creates separate copies that must be maintained independently.

  73. Question 73Build and configure AI agents and workflows

    Which practice keeps an agent's authored topics maintainable as the number of topics grows?

    1. Using descriptive topic names and documented trigger phrases Correct answer.
    2. Disabling the test pane to avoid stale results
    3. Placing all logic in a single large topic
    4. Relying only on generative answers for every request

    Explanation

    Clear names and documented trigger phrases make topics discoverable and reduce overlap as the agent scales. Consolidating everything into one topic or depending solely on generative answers makes behavior harder to reason about.

  74. Question 74Build and configure AI agents and workflows

    A maker wants the agent to greet returning users by name using a value retrieved from a customer system at the start of the conversation. Which approach fits best?

    1. Hard-code a greeting with a placeholder name
    2. Put the customer lookup in the fallback topic
    3. Call the customer system in the conversation-start topic and insert the returned value into the greeting Correct answer.
    4. Ask the user to type their name in every session

    Explanation

    The conversation-start system topic runs first, so a lookup performed there can supply a personalized value for the greeting. Hard-coding or re-asking the user does not use the retrieved data.

  75. Question 75Build and configure AI agents and workflows

    A topic must ask the user for a date and reject responses that are not valid dates before continuing. Which design achieves this?

    1. Redirect to the fallback topic after the first attempt
    2. Use an entity to parse the input and add a condition that re-prompts on failure Correct answer.
    3. Store the raw text and continue regardless of format
    4. Add a message node that lists acceptable date formats

    Explanation

    Entities parse user input into a typed value, and a condition can detect failed parsing and loop back to re-ask. Merely listing formats or storing raw text does not enforce validity.

  76. Question 76Build and configure AI agents and workflows

    A maker wants a topic to end politely and clear the collected variables so a later conversation starts fresh. Which combination is appropriate?

    1. An adaptive card node followed by a set variable node
    2. A condition node followed by a redirect node
    3. A message node followed by a clear variable node and an end conversation node Correct answer.
    4. A transfer node followed by a message node

    Explanation

    Clearing variables before ending ensures stale values do not leak into later conversations, and the end conversation node closes the topic cleanly. The other combinations either skip clearing or do not end the topic.

  77. Question 77Build and configure AI agents and workflows

    A maker wants to test a single topic in isolation without walking through the entire agent conversation. Which capability supports this?

    1. Disabling generative answers globally
    2. Using the topic's own test or preview option Correct answer.
    3. Publishing the agent to a test channel
    4. Deleting unrelated topics before testing

    Explanation

    Topic-level testing lets a maker exercise one topic directly, which speeds iteration and isolates defects. Publishing or deleting topics is disruptive and unnecessary for isolated testing.

  78. Question 78Build and configure AI agents and workflows

    A maker needs an agent to look up a customer's order status by calling an external service during a conversation. Which configuration enables this?

    1. Creating a trigger phrase for each possible order
    2. Adding the service as a knowledge source
    3. Adding the order status to the system message
    4. Registering the service as a tool or action the agent can invoke Correct answer.

    Explanation

    Tools and actions let the agent invoke external services at runtime and use the returned data in the conversation. Knowledge sources ground answers in content but do not execute service calls.

  79. Question 79Build and configure AI agents and workflows

    An agent must call a tool that requires the signed-in user's own credentials rather than a shared service account. Which configuration supports this?

    1. Enable anonymous access on the channel
    2. Embed the credentials in a topic message
    3. Configure per-user authentication for the tool Correct answer.
    4. Store a shared secret in the tool definition

    Explanation

    Per-user authentication prompts each user to sign in so the tool call runs under that user's identity and permissions. Shared secrets and embedded credentials violate least privilege and do not reflect the user.

  80. Question 80Build and configure AI agents and workflows

    A maker wants the agent to confirm a booking by reading back the collected details before finalizing. Which design best achieves this?

    1. A message node that repeats the variables, then a condition that proceeds only on confirmation Correct answer.
    2. A condition node that checks whether the variables are empty
    3. A redirect node to the fallback topic
    4. A set variable node that clears the details

    Explanation

    Reading back the collected values and gating the next step on an explicit confirmation ensures the user verifies the data before it is committed. The other designs neither present the details nor gate on confirmation.

  81. Question 81Build and configure AI agents and workflows

    A maker wants the agent to offer quick-reply buttons for three options while still accepting typed answers. Which node should be used?

    1. A message node with plain text only
    2. A question node configured with choice options Correct answer.
    3. A set variable node
    4. An end conversation node

    Explanation

    A question node with choice options renders selectable buttons and still accepts typed input mapped to the choices. A plain message node cannot present selectable options.

  82. Question 82Build and configure AI agents and workflows

    A maker needs a topic to run a short calculation on two collected numbers and store the result. Which approach is appropriate?

    1. Use a message node with the numbers concatenated
    2. Use an expression in a set variable node to compute the result Correct answer.
    3. Use a condition node to compare the numbers
    4. Use an adaptive card to display the numbers

    Explanation

    Expressions can perform arithmetic on variables, and a set variable node stores the computed result for later use. Concatenation or display does not compute a numeric result.

  83. Question 83Build and configure AI agents and workflows

    A maker wants to prevent a topic from being selected when a user's message contains a specific phrase that belongs to another topic. Which configuration most directly reduces the collision?

    1. Enable generative answers for the first topic
    2. Increase the number of trigger phrases in the first topic
    3. Add the phrase to both topics' trigger phrases
    4. Remove the phrase from the first topic's trigger phrases Correct answer.

    Explanation

    Trigger phrases drive topic matching, so removing a phrase that belongs to another topic reduces misrouting. Adding the phrase to both topics increases ambiguity rather than resolving it.

  84. Question 84Build and configure AI agents and workflows

    A maker wants the agent to respond in the user's preferred language when that preference is known from a prior lookup. Which configuration is appropriate?

    1. Set the language in the system message statically
    2. Add a trigger phrase for each language
    3. Create a separate agent per language
    4. Store the preference in a variable and reference it in responses Correct answer.

    Explanation

    Storing the preference in a variable lets responses reference it and adapt per user, while a static system message applies the same language to everyone. Separate agents or trigger phrases do not dynamically personalize responses.

  85. Question 85Build and configure AI agents and workflows

    An agent must collect a phone number and then call a verification service, retrying up to three times if verification fails. Which design best implements the retry limit?

    1. A redirect to the fallback topic after the first failure
    2. A condition node that loops back to the question and increments a counter variable Correct answer.
    3. An end conversation node after the first failure
    4. A message node that tells the user to try again later

    Explanation

    A counter variable incremented on each failure lets a condition loop back to the question until the limit is reached, after which the flow exits. The other options end or defer instead of retrying with a bound.

  86. Question 86Build and configure AI agents and workflows

    A maker observes that a topic sometimes skips a required question because a variable already holds a value from an earlier topic. Which change best ensures the question is always asked?

    1. Enable generative answers for the topic
    2. Add more trigger phrases to the topic
    3. Move the question to the fallback topic
    4. Clear the relevant variable at the start of the topic Correct answer.

    Explanation

    Variables persist across topics, so a leftover value can cause a question node to be treated as already answered. Clearing the variable at the start of the topic forces the question to be asked again.

  87. Question 87Build and configure AI agents and workflows

    A maker needs an agent to summarize a long user message before deciding which topic to route to. Which approach is most appropriate?

    1. Add a summarization step in a topic that runs before routing decisions Correct answer.
    2. Rely on trigger phrases to match the full message
    3. Increase the number of topics to cover more cases
    4. Store the message in a variable and end the conversation

    Explanation

    Preprocessing the input in a topic before routing lets the agent reduce the message to its intent and make a better routing decision. Trigger phrases match surface text and do not summarize.

  88. Question 88Build and configure AI agents and workflows

    A maker wants to ensure that a tool call made during a conversation does not expose sensitive fields returned by the service. Which approach is best?

    1. Return the full service response to the user
    2. Map only the needed fields into the response and discard the rest Correct answer.
    3. Store the full response in a variable and display it
    4. Log the full response in a message node

    Explanation

    Mapping only the required fields limits what the agent can surface and keeps sensitive data out of the conversation. Returning, logging, or displaying the full response risks exposing sensitive fields.

  89. Question 89Build and configure AI agents and workflows

    A maker needs a topic to branch based on whether a tool call succeeded or returned an error. Which design best supports this?

    1. Redirect to the fallback topic unconditionally
    2. Assume success and continue to the next message
    3. End the conversation after the tool call
    4. Check the tool call result in a condition node and branch accordingly Correct answer.

    Explanation

    Condition nodes can evaluate the result of a tool call and branch to an error path, which is how failures are handled explicitly. Assuming success or ending unconditionally ignores errors.

  90. Question 90Build and configure AI agents and workflows

    A maker wants an agent to answer questions about an internal policy while ensuring the answer always reflects the latest approved version. Which configuration is most appropriate?

    1. Add a trigger phrase for each policy question
    2. Rely on the model's general knowledge of policies
    3. Paste the policy text into the system message
    4. Connect the approved policy document as a knowledge source and keep it current Correct answer.

    Explanation

    A knowledge source grounds answers in the approved document, and keeping that source current ensures answers reflect the latest version. Pasting text or relying on general knowledge risks stale or inaccurate answers.

  91. Question 91Integrate AI solutions with data and services

    An application must send a natural-language prompt to an Azure OpenAI model and receive a completion. Which Azure OpenAI REST endpoint pattern is used for chat completions?

    1. /openai/models/{model-id}/chat
    2. /openai/completions/{deployment-id}/chat
    3. /openai/chat/{deployment-id}/completions
    4. /openai/deployments/{deployment-id}/chat/completions Correct answer.

    Explanation

    Azure OpenAI routes requests through a deployment name rather than a model name, and chat completions are addressed as /openai/deployments/{deployment-id}/chat/completions on the resource endpoint. The deployment identifier in the path is what binds the call to a specific model configuration.

  92. Question 92Integrate AI solutions with data and services

    A solution stores the Azure OpenAI API key in Azure Key Vault and the app runs in Azure App Service. Which approach lets the app read the secret without embedding credentials in code or configuration?

    1. Expose the vault publicly and authenticate with a shared access signature
    2. Store the key in an app setting named after the vault and read it at startup
    3. Assign a managed identity to the App Service and grant it Key Vault secret read access Correct answer.
    4. Download the vault to local storage during deployment

    Explanation

    Managed identities let the platform issue and rotate credentials for the workload, and Key Vault access policies or RBAC then authorize that identity to read secrets. This removes long-lived secrets from code and configuration entirely.

  93. Question 93Integrate AI solutions with data and services

    Which Azure service is designed to index documents and support full-text and vector search queries for a retrieval-augmented generation pipeline?

    1. Azure Cosmos DB for NoSQL
    2. Azure Data Factory
    3. Azure Service Bus
    4. Azure AI Search Correct answer.

    Explanation

    Azure AI Search provides indexes with text and vector fields plus hybrid ranking, which is what RAG retrieval needs. The other services store or move data but do not provide a search index with relevance ranking.

  94. Question 94Integrate AI solutions with data and services

    An app calls an Azure OpenAI model and receives HTTP 429. Which action best addresses the root cause?

    1. Disable content filtering on the deployment
    2. Switch the request to a GET method
    3. Implement retry with exponential backoff and request quota or throughput increases Correct answer.
    4. Increase the max_tokens value in the request

    Explanation

    HTTP 429 indicates rate limiting or quota exhaustion, so the client should back off and retry while capacity is adjusted. Neither request shape nor content settings change the throttling condition.

  95. Question 95Integrate AI solutions with data and services

    What is the primary purpose of a system message in a chat completions request?

    1. To control the HTTP status code returned
    2. To store the user's previous turns for auditing
    3. To specify the model deployment name
    4. To set persistent instructions and behavior for the assistant Correct answer.

    Explanation

    The system role carries instructions that condition the assistant across the conversation. User and assistant roles carry the dialogue itself, not the behavioral framing.

  96. Question 96Integrate AI solutions with data and services

    A team needs the model to return data that conforms to a specific JSON schema so downstream code can parse it reliably. Which capability should they use?

    1. Structured outputs with a JSON schema Correct answer.
    2. A higher temperature setting
    3. Streaming responses
    4. A larger context window

    Explanation

    Structured outputs constrain generation to a supplied schema, which makes parsing deterministic. Sampling settings and context size affect content or capacity but do not enforce shape.

  97. Question 97Integrate AI solutions with data and services

    Which embedding input produces a vector suitable for semantic similarity comparison with another embedding?

    1. A hash of the text chunk
    2. The raw text chunk encoded with the same embedding model Correct answer.
    3. The document filename
    4. The token count of the text chunk

    Explanation

    Semantic similarity requires both texts to be embedded by the same model into the same vector space. Hashes, counts, and filenames carry no semantic signal comparable across texts.

  98. Question 98Integrate AI solutions with data and services

    A function-calling flow returns a tool call from the model. What must the application do before the model can produce a final natural-language answer?

    1. Discard the tool call and resend the original prompt
    2. Convert the tool call into a system message
    3. Execute the function and send the result back as a tool message Correct answer.
    4. Increase the temperature and retry

    Explanation

    The model only proposes a call; the application executes it and returns the output in a tool-role message so the model can incorporate the result. Without that result the model lacks the information it requested.

  99. Question 99Integrate AI solutions with data and services

    A RAG pipeline splits a long policy document into chunks. Which chunking strategy best preserves answerable context across chunk boundaries?

    1. Split into one chunk per page regardless of content
    2. Split every 50 characters with no overlap
    3. Split into chunks of exactly one sentence each
    4. Split on semantic sections with a modest overlap between adjacent chunks Correct answer.

    Explanation

    Section-aware splitting keeps related content together, and overlap reduces the chance that an answer is cut across a boundary. Arbitrary or overly small splits fragment meaning and hurt retrieval quality.

  100. Question 100Integrate AI solutions with data and services

    An index contains both a text field and a vector field for the same content. A query supplies a search string and a query vector. What does hybrid search with reciprocal rank fusion do?

    1. Returns the union of results in insertion order
    2. Runs only the vector query and ignores the text query
    3. Combines the ranked results from the text and vector queries into a single ranking Correct answer.
    4. Averages the raw scores of both queries without ranking

    Explanation

    Hybrid search issues both queries and fuses their rankings, commonly with reciprocal rank fusion, which is robust to differing score scales. This improves recall over either modality alone.

  101. Question 101Integrate AI solutions with data and services

    A retrieval step returns many low-relevance chunks that inflate prompt size and cost. Which change most directly improves precision before the chunks reach the model?

    1. Increase the model temperature
    2. Remove the vector field from the index
    3. Increase top-k to return more chunks
    4. Apply a semantic reranker and raise the relevance threshold Correct answer.

    Explanation

    Reranking rescores candidates for query relevance and a threshold filters out weak matches, reducing prompt bloat. Returning more candidates or removing the vector field works against precision.

  102. Question 102Integrate AI solutions with data and services

    An application must summarize a document that exceeds the model's context window. Which approach is appropriate?

    1. Send the document as the system message
    2. Map-reduce summarization over chunks, then combine partial summaries Correct answer.
    3. Truncate the document to the first page
    4. Raise max_tokens beyond the context limit

    Explanation

    Chunked map-reduce summarization processes each part within limits and then synthesizes the partial results. Truncation loses content and max_tokens cannot exceed the model's context capacity.

  103. Question 103Integrate AI solutions with data and services

    Which practice best protects against prompt injection when an agent reads untrusted web content and also has tool access?

    1. Give the agent broader tool permissions to handle edge cases
    2. Put the untrusted content in the system message
    3. Disable logging to avoid leaking the injected text
    4. Treat retrieved content as untrusted data and gate tool calls behind explicit authorization Correct answer.

    Explanation

    Content fetched from external sources can contain instructions that try to hijack the agent, so it should be handled as data and tool invocations should be authorized independently. Elevating untrusted text or widening permissions increases the attack surface.

  104. Question 104Integrate AI solutions with data and services

    An app needs deterministic, low-variance classification output from a chat model. Which parameter setting best supports that goal?

    1. A high top_p such as 1.0 combined with high temperature
    2. A low temperature such as 0 Correct answer.
    3. A large max_tokens value
    4. A high temperature such as 1.5

    Explanation

    Lower temperature narrows sampling so the model favors its highest-probability tokens, which increases consistency for classification. High sampling settings increase variation, and max_tokens only bounds length.

  105. Question 105Integrate AI solutions with data and services

    A pipeline must generate embeddings for millions of documents. Which design most improves throughput and resilience?

    1. Batch requests within token limits and process with parallel workers that retry failures Correct answer.
    2. Reduce the embedding dimensions by truncating the vectors
    3. Send one request per document sequentially
    4. Concatenate all documents into a single request

    Explanation

    Batching amortizes request overhead and parallel workers with retries sustain throughput under transient failures. Sequential single calls are slow, and a single giant request exceeds limits.

  106. Question 106Integrate AI solutions with data and services

    A solution uses Azure AI Document Intelligence to extract tables from PDFs before indexing. Why is this preferable to sending raw PDF bytes to the model for extraction?

    1. It encrypts the PDF automatically
    2. It guarantees the model will never hallucinate
    3. It produces structured, layout-aware output that is cheaper and more reliable to index Correct answer.
    4. It removes the need for any indexing step

    Explanation

    A dedicated extraction service returns structured fields and tables with layout awareness, which is more reliable and cost-efficient than paying a model to interpret raw bytes. Indexing is still required to make the content retrievable.

  107. Question 107Integrate AI solutions with data and services

    An application must call a model with a private, organization-specific endpoint that keeps data within a defined network boundary. Which configuration supports this?

    1. A private endpoint on the Azure OpenAI resource accessed through a virtual network Correct answer.
    2. A proxy that forwards requests to a public endpoint
    3. A shared public endpoint used by all tenants
    4. A public endpoint with the key appended to the URL

    Explanation

    Private endpoints attach the resource to a virtual network so traffic stays on the private network path. Public or shared endpoints and forwarding proxies route traffic outside the boundary.

  108. Question 108Integrate AI solutions with data and services

    A team wants reproducible evaluation of a RAG system's answer quality across prompt changes. Which approach is most sound?

    1. Use a fixed labeled question set and compare groundedness and relevance metrics across runs Correct answer.
    2. Eyeball a few answers after each change
    3. Ask the model whether its own answers are good without a reference set
    4. Only measure latency and token cost

    Explanation

    A fixed labeled set makes comparisons across prompt variants meaningful, and groundedness and relevance metrics quantify quality. Anecdotal review or self-assessment without references is not reproducible.

  109. Question 109Integrate AI solutions with data and services

    A function-calling schema defines a parameter as an enum of allowed values. What is the main benefit?

    1. It increases the model's context window
    2. It encrypts the parameter value
    3. It bypasses content filtering
    4. It constrains the model to emit valid values that the function can accept Correct answer.

    Explanation

    Enumerated schemas tell the model the permitted values, reducing invalid arguments that would fail validation. Schema constraints do not affect encryption, context size, or filtering.

  110. Question 110Integrate AI solutions with data and services

    Which Azure service combination best supports an event-driven pipeline that ingests a new document and triggers indexing without polling?

    1. A message queue polled once per day
    2. Blob storage events routed through Event Grid to a function that indexes the document Correct answer.
    3. A manual upload step followed by a script
    4. A scheduled job that scans storage every hour

    Explanation

    Storage events delivered through Event Grid push notifications to a handler, so indexing begins as soon as a document arrives. Scheduled scans and polling add latency and unnecessary work.

  111. Question 111Integrate AI solutions with data and services

    An agent must answer questions over a large corpus and also take actions such as creating tickets. Which architecture best separates these concerns?

    1. Disable retrieval and rely on the model's training data
    2. Retrieve grounding context first, then let the model choose among authorized tools Correct answer.
    3. Let the model call any API it discovers at runtime
    4. Embed the entire corpus in the system prompt

    Explanation

    Separating retrieval from action keeps grounding focused and confines tool use to an authorized set, which limits risk. Embedding the whole corpus or granting open API access is infeasible and unsafe.

  112. Question 112Integrate AI solutions with data and services

    A RAG system returns confident but incorrect answers for questions whose answers span two distant sections of a document. Which change most directly addresses this failure?

    1. Reduce the model temperature to zero
    2. Shorten the system prompt
    3. Increase chunk overlap and retrieve multiple chunks before generating Correct answer.
    4. Lower the relevance threshold to admit more chunks

    Explanation

    Answers spanning distant sections require that both pieces be retrieved and presented together, which overlap and multi-chunk retrieval support. Temperature and prompt length do not fix missing context.

  113. Question 113Integrate AI solutions with data and services

    An application streams completions to users. Which design consideration is most important for handling a mid-stream failure?

    1. Increase max_tokens to prevent failures
    2. Buffer the entire response before displaying anything
    3. Detect the broken stream and surface a recoverable error or resume strategy to the client Correct answer.
    4. Disable retries for all requests

    Explanation

    Streaming delivers partial output, so a mid-stream failure leaves the client with incomplete content that must be detected and handled, for example by erroring or resuming. Buffering defeats streaming and max_tokens does not prevent network failures.

  114. Question 114Integrate AI solutions with data and services

    A team wants to reduce token cost for a high-volume classification workload without losing accuracy on hard cases. Which strategy is most effective?

    1. Send every request to the largest model with maximum max_tokens
    2. Increase temperature so answers vary
    3. Remove all few-shot examples to shorten prompts
    4. Route easy cases to a smaller model and escalate uncertain cases to a larger model Correct answer.

    Explanation

    Model routing sends the bulk of simple traffic to a cheaper model while preserving quality on difficult inputs by escalating them. Using the largest model everywhere or removing guidance raises cost or hurts accuracy.

  115. Question 115Integrate AI solutions with data and services

    An index must support filtering results by tenant before vector similarity is computed. Which index design achieves correct and efficient tenant isolation?

    1. Rely on the model to ignore other tenants' content
    2. Encode the tenant in the vector values
    3. Add a tenant identifier field marked filterable and apply a filter on every query Correct answer.
    4. Store each tenant's vectors in the same field without a tenant attribute

    Explanation

    A filterable tenant field lets the query restrict the candidate set before ranking, which enforces isolation and improves efficiency. Without an attribute, results can cross tenants.

  116. Question 116Integrate AI solutions with data and services

    A developer wants the model to call an external weather API through a tool. Which component defines the tool's name, purpose, and input parameters so the model can decide when and how to invoke it?

    1. The temperature parameter
    2. The system message
    3. The tool definition schema Correct answer.
    4. The completion's finish reason

    Explanation

    A tool definition is a structured description, typically JSON Schema, that tells the model the tool's name, what it does, and what arguments it accepts. The model uses this schema to decide whether to emit a tool call and to construct valid arguments. System messages can give general instructions but do not formally declare callable tools.

  117. Question 117Integrate AI solutions with data and services

    An application receives a tool call from the model requesting a customer record by ID. Which step must occur before the model can incorporate the tool's result into its answer?

    1. The application must convert the tool result into an embedding and store it in an index.
    2. The application must lower the temperature to zero and resend the original prompt.
    3. The application must execute the tool and append a tool-result message to the conversation. Correct answer.
    4. The application must retrain the model on the tool output.

    Explanation

    The model only proposes a tool call; the host application is responsible for running the function and returning its output as a tool message in the conversation. The model then continues generation with that result in context. Retraining, temperature changes, or embedding storage are not part of the standard tool-call round trip.

  118. Question 118Integrate AI solutions with data and services

    A team stores embeddings in a vector index and wants to retrieve only documents belonging to the current user. Which index design enforces this constraint correctly?

    1. Rely on the LLM to refuse answers about other users after retrieval.
    2. Store the user ID as a filterable metadata field on each vector and apply an equality filter on it during query. Correct answer.
    3. Store the user ID inside the embedding vector as an extra dimension.
    4. Encode the user ID in the document filename and rely on the model to ignore other users' documents.

    Explanation

    Tenant or user isolation must be enforced by the query layer, typically through a filterable metadata field applied during vector search. This restricts the candidate set before similarity ranking. Encoding identity in vectors, filenames, or relying on the model are not reliable access controls.

  119. Question 119Integrate AI solutions with data and services

    A developer needs to send a long conversation to a chat model while preserving the most recent turns. Which approach best manages the context window?

    1. Increase the model's temperature to compress the history.
    2. Summarize older turns and keep recent turns verbatim within the token budget. Correct answer.
    3. Send the entire conversation every time regardless of length.
    4. Remove all prior turns and send only the latest user message.

    Explanation

    Context windows are finite, so applications typically preserve recent turns and condense older history into a summary that fits the budget. This retains relevant context while avoiding overflow. Sending everything can exceed limits, and dropping all history loses continuity.

  120. Question 120Integrate AI solutions with data and services

    An application needs to extract structured fields from scanned invoices and also needs the model to reason over the extracted values. Which combination is most appropriate?

    1. Use a text embedding model alone to produce the invoice fields.
    2. Use a vector index alone to retrieve invoice fields.
    3. Use a document extraction service to produce structured fields, then pass those fields to a chat model for reasoning. Correct answer.
    4. Use a chat model alone on the raw image without any extraction step.

    Explanation

    Document extraction services are designed to pull structured fields from scanned forms, while chat models excel at reasoning over structured input. Combining them separates reliable extraction from interpretation. Embeddings and indexes support retrieval, not field extraction.

  121. Question 121Integrate AI solutions with data and services

    A developer wants the model to decide between two tools: one for looking up orders and one for issuing refunds. Which design most improves the model's tool-selection accuracy?

    1. Merge both tools into a single tool with no parameters.
    2. Give both tools identical descriptions and parameter names.
    3. Provide distinct, specific descriptions of each tool's purpose and required inputs. Correct answer.
    4. Set the model's temperature to the maximum value.

    Explanation

    Tool selection depends heavily on the descriptions supplied in the tool definitions. Clear, distinct descriptions of purpose and inputs help the model choose the right tool. Ambiguous or identical descriptions increase misselection.

  122. Question 122Integrate AI solutions with data and services

    An agent must call a tool that occasionally returns transient errors. Which application-level behavior most improves reliability?

    1. Ignore the error and fabricate a plausible tool result.
    2. Disable the tool permanently after the first error.
    3. Retry indefinitely until the tool succeeds.
    4. Retry the tool call with exponential backoff and a bounded attempt limit. Correct answer.

    Explanation

    Transient failures are best handled with bounded retries using exponential backoff, which avoids overwhelming the dependency while recovering from temporary faults. Unbounded retries can hang the workflow, and fabricating results corrupts the model's context.

  123. Question 123Integrate AI solutions with data and services

    A retrieval-augmented application must cite the source document for each claim. Which design most directly supports accurate citations?

    1. Increase the model's temperature so citations vary.
    2. Store citations only in the vector index and never send them to the model.
    3. Include document identifiers and source text in the prompt context and require the model to reference them. Correct answer.
    4. Ask the model to invent plausible source names when none are provided.

    Explanation

    For the model to cite sources, the retrieved passages must carry identifiers and be present in the prompt, and the instructions must require referencing them. The model cannot cite material it never receives.

  124. Question 124Integrate AI solutions with data and services

    An application calls an Azure OpenAI model and must ensure requests are authenticated without embedding long-lived keys in client code. Which approach is appropriate?

    1. Disable authentication on the endpoint.
    2. Hard-code the API key in the mobile app.
    3. Use Microsoft Entra ID authentication with managed identities or scoped tokens. Correct answer.
    4. Email the API key to each user.

    Explanation

    Microsoft Entra ID with managed identities or short-lived tokens avoids embedding static secrets in client code and supports least-privilege access. Hard-coded keys and unauthenticated endpoints create serious exposure.

  125. Question 125Integrate AI solutions with data and services

    A developer wants to log model interactions for auditing while minimizing exposure of sensitive content. Which practice is most appropriate?

    1. Log full prompts and responses including all personal data indefinitely.
    2. Store logs in a public blob container for easy access.
    3. Redact or hash sensitive fields before logging and apply retention policies. Correct answer.
    4. Disable logging entirely to avoid any risk.

    Explanation

    Auditing requires logs, but sensitive fields should be redacted or hashed and retention should be governed by policy. This balances traceability with privacy and compliance.

  126. Question 126Integrate AI solutions with data and services

    A team needs the model to answer questions using data from an internal SQL database that changes frequently. Which approach keeps answers current?

    1. Expose a read-only query tool the model can call at request time. Correct answer.
    2. Embed the entire database once and never refresh it.
    3. Cache answers for a year to reduce calls.
    4. Fine-tune the model on a nightly export of the database.

    Explanation

    A read-only query tool lets the model retrieve current data at request time rather than relying on stale training or cached content. Fine-tuning on exports and one-time embeddings quickly become outdated.

  127. Question 127Integrate AI solutions with data and services

    A developer wants to reduce hallucinations in a retrieval-augmented system when the retrieved context does not contain the answer. Which design is most appropriate?

    1. Raise the temperature so the model is more creative.
    2. Always return the top-scoring chunk as the answer verbatim.
    3. Instruct the model to state that the answer is unavailable when the context lacks it. Correct answer.
    4. Remove the retrieved context to simplify the prompt.

    Explanation

    Grounding instructions that permit an explicit 'not found' response reduce fabricated answers when context is insufficient. Removing context or forcing a chunk as the answer increases incorrect output.

  128. Question 128Integrate AI solutions with data and services

    An application must process a batch of documents through a model and report per-document status. Which design best supports observability and partial failure?

    1. Suppress all errors and report success for every document.
    2. Process documents individually or in small groups and record status and errors per item. Correct answer.
    3. Retry the entire batch indefinitely on any failure.
    4. Process all documents in one request and fail the whole batch on any error.

    Explanation

    Per-item or small-group processing isolates failures and enables accurate status reporting. Whole-batch failure handling and suppressed errors obscure which documents succeeded.

  129. Question 129Integrate AI solutions with data and services

    A team wants the model to summarize a table returned by a database tool. Which input format best preserves the table's structure for the model?

    1. A structured representation such as JSON with column names and rows. Correct answer.
    2. A random sample of individual cell values.
    3. A single comma-separated string with no headers.
    4. A base64-encoded binary dump of the table.

    Explanation

    Structured formats like JSON preserve column names and row relationships, making it easier for the model to interpret and summarize accurately. Unlabeled strings and binary dumps lose structure.

  130. Question 130Integrate AI solutions with data and services

    A developer is choosing between a small model and a large model for a tool-calling workflow that must handle ambiguous user requests. Which factor most directly justifies using the larger model?

    1. The larger model generally follows complex instructions and selects tools more reliably. Correct answer.
    2. The larger model always has a lower per-token price.
    3. The larger model eliminates the need for tool definitions.
    4. The larger model removes the need for application-side validation.

    Explanation

    Larger models typically handle complex instructions and ambiguous tool-selection scenarios more reliably, which matters when requests are unclear. They do not remove the need for tool schemas or validation, and they are usually not cheaper.

  131. Question 131Test, monitor, and optimize AI solutions

    Which metric best measures the percentage of relevant items that a retrieval system successfully returned?

    1. Perplexity
    2. Throughput
    3. Latency
    4. Recall Correct answer.

    Explanation

    Recall measures the proportion of all relevant items in the corpus that the system actually retrieved. Latency and throughput describe performance, not retrieval completeness, and perplexity measures language-model prediction uncertainty.

  132. Question 132Test, monitor, and optimize AI solutions

    A team wants a single number summarizing how well a retrieval system ranks relevant documents near the top. Which metric should they use?

    1. F1 of the classifier head
    2. Token count
    3. Root mean squared error
    4. Mean reciprocal rank Correct answer.

    Explanation

    Mean reciprocal rank averages the inverse rank of the first relevant result, rewarding systems that place relevant items early. RMSE and token count do not capture ranking position, and classifier F1 is unrelated to retrieval ordering.

  133. Question 133Test, monitor, and optimize AI solutions

    An evaluation set is built by having a model generate answers and a second model score them. What is the primary risk of this approach?

    1. Latency becomes deterministic
    2. Judge bias may inflate scores Correct answer.
    3. Token costs drop to zero
    4. Ground truth becomes exhaustive

    Explanation

    Model-based judges can carry position, verbosity, or self-preference biases that skew scores. Cost and latency are operational concerns, not accuracy risks, and model scoring does not make ground truth more complete.

  134. Question 134Test, monitor, and optimize AI solutions

    Production traffic shows a RAG assistant answering correctly for common questions but failing on rare product codes. Which evaluation gap does this most likely reveal?

    1. Embedding dimensions are too large
    2. Evaluation set lacks long-tail coverage Correct answer.
    3. Temperature is set too low
    4. Evaluation set has too many hard cases

    Explanation

    Failures concentrated on rare inputs usually mean the evaluation set underrepresents the long tail, so the weakness was never measured. Temperature and embedding size are configuration choices that do not by themselves explain a coverage gap.

  135. Question 135Test, monitor, and optimize AI solutions

    Which practice most directly reduces overfitting of prompt tuning to a specific evaluation set?

    1. Holding out a separate validation set Correct answer.
    2. Increasing the model temperature
    3. Adding more examples to the tuning set only
    4. Shortening the system prompt

    Explanation

    A held-out validation set is not used during prompt iteration, so it gives an unbiased estimate of generalization. Adding more tuning examples or adjusting temperature does not address leakage into the tuning set.

  136. Question 136Test, monitor, and optimize AI solutions

    What is the main purpose of a regression test suite for an AI feature?

    1. Replace human review entirely
    2. Guarantee zero hallucinations
    3. Reduce model parameter count
    4. Detect quality drops after changes Correct answer.

    Explanation

    Regression suites rerun fixed cases after prompt, model, or data changes to catch quality degradation. They cannot guarantee zero hallucinations or replace human judgment, and they do not change model size.

  137. Question 137Test, monitor, and optimize AI solutions

    A team needs to detect when input data distribution shifts in production. Which monitoring approach is most appropriate?

    1. Track feature and embedding drift statistics Correct answer.
    2. Monitor only GPU utilization
    3. Review the changelog weekly
    4. Count the number of deployments

    Explanation

    Drift detection compares live feature or embedding distributions against a reference baseline to flag distribution shift. Infrastructure metrics and deployment counts say nothing about input distribution.

  138. Question 138Test, monitor, and optimize AI solutions

    Which signal most directly indicates that a deployed model's output quality is degrading even though infrastructure is healthy?

    1. Unchanged container image
    2. Rising rate of low-confidence or flagged outputs Correct answer.
    3. Stable CPU usage
    4. Constant request volume

    Explanation

    Quality degradation often surfaces as more low-confidence outputs or more items flagged by guardrails, even when servers are fine. Stable infrastructure metrics confirm only that the system is running, not that it is accurate.

  139. Question 139Test, monitor, and optimize AI solutions

    Which practice best supports reproducibility of an evaluation run?

    1. Rotate the evaluation set each run
    2. Disable logging to save storage
    3. Rely on live production traffic
    4. Record model version, prompt, and data snapshot Correct answer.

    Explanation

    Reproducibility requires capturing the exact model version, prompt, parameters, and dataset used so a run can be replayed. Live traffic and rotating datasets change inputs, and disabling logging removes the evidence needed.

  140. Question 140Test, monitor, and optimize AI solutions

    A RAG pipeline returns answers that cite documents which do not actually support the claim. Which metric should be added to the evaluation?

    1. Cold-start latency
    2. Index build time
    3. Token throughput
    4. Groundedness or faithfulness score Correct answer.

    Explanation

    Groundedness or faithfulness measures whether each claim is supported by the retrieved source, directly targeting unsupported citations. Latency and build-time metrics are operational and say nothing about claim support.

  141. Question 141Test, monitor, and optimize AI solutions

    Which approach most effectively reduces hallucinations in a retrieval-augmented assistant?

    1. Require answers to cite retrieved passages Correct answer.
    2. Increase the model temperature
    3. Shorten the user query
    4. Remove the retrieval step

    Explanation

    Grounding answers in retrieved passages and requiring citations constrains the model to supported content, reducing fabrication. Raising temperature increases randomness, and removing retrieval eliminates the grounding source.

  142. Question 142Test, monitor, and optimize AI solutions

    What does a precision metric of 0.9 for a retrieval system indicate?

    1. The index holds 90 percent of documents
    2. Nine of ten returned items are relevant Correct answer.
    3. The system is 90 percent faster
    4. Nine of ten relevant items were found

    Explanation

    Precision is the fraction of retrieved items that are relevant, so 0.9 means nine in ten returned items are relevant. Recall, not precision, measures the fraction of relevant items found.

  143. Question 143Test, monitor, and optimize AI solutions

    A team compares two prompt variants and finds variant A scores higher on average but has much higher variance across runs. What should they consider before shipping A?

    1. Reduce the evaluation set size
    2. Run repeated trials and compare distributions Correct answer.
    3. Ignore variance as noise
    4. Ship A because the mean is higher

    Explanation

    High variance means the mean may be unstable, so repeated trials and distribution comparison are needed before concluding A is better. Shipping on a single mean ignores run-to-run variability, and shrinking the set worsens estimates.

  144. Question 144Test, monitor, and optimize AI solutions

    Which metric is most appropriate for evaluating a summarization model against human-written reference summaries?

    1. Embedding dimensionality
    2. Mean time to first token
    3. ROUGE overlap score Correct answer.
    4. Index recall

    Explanation

    ROUGE measures n-gram overlap between generated and reference summaries, a standard reference-based summarization metric. Latency, embedding size, and index recall do not measure summary quality.

  145. Question 145Test, monitor, and optimize AI solutions

    A team wants to catch prompt injection attempts in user input before they reach the model. Which control is most direct?

    1. Increasing max tokens
    2. Lowering the model temperature
    3. Input filtering with injection classifiers Correct answer.
    4. Caching responses

    Explanation

    Input filtering that detects injection patterns or uses a classifier intercepts malicious prompts before inference. Temperature, token limits, and caching affect generation behavior but do not screen for injection.

  146. Question 146Test, monitor, and optimize AI solutions

    Which practice best protects sensitive data when logging AI interactions for monitoring?

    1. Share raw logs across all teams
    2. Redact or hash personally identifiable fields Correct answer.
    3. Store logs in plaintext indefinitely
    4. Log full prompts with no filtering

    Explanation

    Redacting or hashing sensitive fields lets teams monitor behavior while limiting exposure of personal data. Logging raw prompts, keeping plaintext indefinitely, or broadly sharing logs increases privacy and compliance risk.

  147. Question 147Test, monitor, and optimize AI solutions

    What is the purpose of a guardrail that blocks unsafe model outputs at inference time?

    1. Prevent harmful content from reaching users Correct answer.
    2. Speed up index rebuilds
    3. Improve model accuracy on benchmarks
    4. Reduce embedding storage costs

    Explanation

    Output guardrails screen generated content and block unsafe responses before users see them. They are a safety control, not an accuracy booster or a cost or indexing optimization.

  148. Question 148Test, monitor, and optimize AI solutions

    An evaluation shows high recall but low precision for a retrieval system. What is the most likely user impact?

    1. The model responds too slowly
    2. Relevant documents are missed
    3. Answers include irrelevant context Correct answer.
    4. The index is too small

    Explanation

    Low precision means many retrieved items are irrelevant, so the model receives noisy context. Low recall, not low precision, would mean relevant documents are missed.

  149. Question 149Test, monitor, and optimize AI solutions

    Which approach best evaluates whether a chatbot's answers are helpful, when no single reference answer exists?

    1. BLEU against one reference
    2. Rubric-based human or model grading Correct answer.
    3. Exact string match
    4. Token count comparison

    Explanation

    When many valid answers exist, rubric-based grading captures helpfulness dimensions that exact match or single-reference overlap cannot. String match and BLEU penalize valid paraphrases.

  150. Question 150Test, monitor, and optimize AI solutions

    A team notices evaluation scores drop only after they upgraded the embedding model. What should they check first?

    1. Whether the index was rebuilt with the new embeddings Correct answer.
    2. Whether the changelog is complete
    3. Whether GPU clocks are stable
    4. Whether the UI changed

    Explanation

    Changing the embedding model requires reindexing documents with the same model, or queries and documents live in mismatched vector spaces. GPU clocks, UI changes, and changelog completeness do not explain embedding-related score drops.

  151. Question 151Test, monitor, and optimize AI solutions

    Which metric best captures whether a classification model's positive predictions are trustworthy?

    1. Accuracy on a balanced set
    2. Recall
    3. Training loss
    4. Precision Correct answer.

    Explanation

    Precision is the fraction of predicted positives that are actually positive, which reflects trustworthiness of positive predictions. Recall measures coverage of actual positives, and training loss is not a deployment quality measure.

  152. Question 152Test, monitor, and optimize AI solutions

    What is the primary benefit of canary deployment for an AI feature?

    1. Guarantee identical outputs
    2. Eliminate the need for evaluation
    3. Remove the need for monitoring
    4. Limit exposure of a new version to a small traffic share Correct answer.

    Explanation

    Canary deployment routes a small fraction of traffic to the new version so problems surface with limited user impact. It does not remove the need for evaluation or monitoring, nor does it guarantee output consistency.

  153. Question 153Test, monitor, and optimize AI solutions

    A team wants to measure end-to-end latency of a RAG request. Which breakdown is most useful for diagnosis?

    1. Only the number of retrieved documents
    2. Only model parameter count
    3. Retrieval, prompt assembly, and generation times separately Correct answer.
    4. Only total wall-clock time

    Explanation

    Splitting latency into retrieval, assembly, and generation stages localizes the bottleneck. A single total, parameter count, or document count alone cannot show which stage is slow.

  154. Question 154Test, monitor, and optimize AI solutions

    An evaluation harness reports identical scores for two very different prompts. What is the most likely explanation to investigate?

    1. The harness may be caching or ignoring the prompt Correct answer.
    2. The prompts are semantically identical
    3. The model is perfectly deterministic
    4. The dataset is too large

    Explanation

    Identical scores from different prompts often indicate the harness is reusing cached results or failing to pass the prompt through. Perfect determinism would not make different prompts score identically, and dataset size does not explain it.

  155. Question 155Test, monitor, and optimize AI solutions

    Which practice best supports continuous evaluation of a deployed AI system?

    1. One-time evaluation before launch
    2. Manual review of a single user session
    3. Disabling metrics collection
    4. Automated evaluation on sampled live traffic Correct answer.

    Explanation

    Sampling live traffic and scoring it automatically provides ongoing visibility into quality after launch. One-time evaluation, single-session review, or disabled metrics leave regressions undetected.

  156. Question 156Test, monitor, and optimize AI solutions

    A team wants a single metric that summarizes how well a RAG assistant's generated answer agrees with the ground-truth answer at the semantic level, even when wording differs. Which metric is most appropriate?

    1. Token-level edit distance to the reference answer
    2. Character count of the generated answer
    3. Exact string match against the reference answer
    4. Semantic similarity scoring between generated and reference answers Correct answer.

    Explanation

    Semantic similarity compares meaning rather than surface form, so paraphrased but correct answers still score highly. Surface metrics such as exact match or edit distance penalize valid rewordings. Answer length carries no signal about correctness.

  157. Question 157Test, monitor, and optimize AI solutions

    An evaluation harness scores a summarization model using ROUGE, and reviewers complain the scores do not reflect obvious factual errors in the summaries. Which change best addresses this gap?

    1. Lower the n-gram order used in the ROUGE calculation
    2. Increase the weight of ROUGE-L over ROUGE-1
    3. Average ROUGE across more reference summaries
    4. Add a faithfulness check that verifies each claim is supported by the source document Correct answer.

    Explanation

    n-gram overlap metrics reward lexical similarity and cannot detect confidently stated falsehoods. A faithfulness check compares generated claims against the source and catches unsupported content. Re-weighting or averaging ROUGE variants keeps the same blind spot.

  158. Question 158Test, monitor, and optimize AI solutions

    A team wants to track whether a deployed intent classifier is drifting in production without waiting for labeled feedback. Which approach is most suitable?

    1. Retrain the classifier every night on the same training data
    2. Increase the temperature of the classifier's softmax output
    3. Monitor the distribution of predicted intent labels and confidence scores over time Correct answer.
    4. Reduce the number of intent classes to the most frequent ones

    Explanation

    Unlabeled production data still yields prediction distributions and confidence values, and shifts in those signals can flag drift before labels arrive. Retraining on unchanged data cannot detect new input patterns. Temperature and class reduction change outputs but do not measure drift.

  159. Question 159Test, monitor, and optimize AI solutions

    A team must decide whether to ship a new prompt template. They run an offline evaluation and plan a staged rollout. Which combination best balances risk and evidence before full release?

    1. Full rollout only after every reviewer manually reads 1,000 outputs
    2. Full rollout immediately after offline evaluation passes
    3. Replacing the previous prompt without keeping a fallback path
    4. Offline evaluation followed by a small-percentage canary with online quality and latency monitoring Correct answer.

    Explanation

    Offline evaluation filters clearly bad candidates, and a canary exposes the change to limited real traffic while online metrics confirm behavior. This limits blast radius and supplies production evidence. Immediate full rollout or removing fallbacks increases risk, and manual review alone does not scale.

  160. Question 160Test, monitor, and optimize AI solutions

    Which metric directly quantifies how often a RAG system retrieves at least one relevant document for each query?

    1. BLEU
    2. Normalized discounted cumulative gain
    3. Recall at k Correct answer.
    4. Mean reciprocal rank

    Explanation

    Recall at k measures the fraction of queries for which relevant documents appear among the top k results, which matches the stated question. MRR and NDCG reward ranking position rather than presence. BLEU is a text-generation overlap metric.

  161. Question 161Test, monitor, and optimize AI solutions

    A monitoring dashboard shows a model's latency and error rate are stable, but user thumbs-down feedback has risen sharply. Which additional measurement most directly explains the cause?

    1. CPU utilization of the inference servers
    2. Count of HTTP 200 responses
    3. Periodic sampling of conversations scored by an LLM judge against a rubric Correct answer.
    4. Number of tokens in the system prompt

    Explanation

    Stable infrastructure metrics rule out availability problems, so the regression is likely in output quality. Sampling and rubric-based judging surfaces quality issues that infrastructure telemetry misses. Server CPU, prompt length, and success response counts do not measure answer quality.

  162. Question 162Test, monitor, and optimize AI solutions

    What is the primary purpose of an offline evaluation set that is held out and never used for prompt tuning?

    1. To reduce inference latency in production
    2. To increase the size of the training corpus
    3. To provide an unbiased estimate of performance on unseen inputs Correct answer.
    4. To replace the need for production monitoring

    Explanation

    A held-out set is not influenced by tuning decisions, so its scores estimate generalization rather than fit to the development data. It does not train the model or affect latency, and it cannot substitute for observing live traffic.

  163. Question 163Test, monitor, and optimize AI solutions

    A team wants to reduce the cost of running an LLM-as-judge over thousands of evaluation cases while keeping scores reliable. Which approach is most appropriate?

    1. Remove the rubric so the judge responds faster
    2. Increase the judge's temperature to diversify its scores
    3. Judge only the longest responses in the set
    4. Use a smaller judge model validated against a human-labeled sample before adopting it Correct answer.

    Explanation

    A cheaper judge is acceptable only if its agreement with human labels is established on a validation sample. Removing the rubric reduces consistency, selective judging biases the sample, and higher temperature adds noise rather than reliability.

  164. Question 164Test, monitor, and optimize AI solutions

    A red-team exercise finds the assistant will reveal its system prompt when asked to repeat everything above. Which mitigation most directly addresses this class of exposure?

    1. Shorten the system prompt to a single sentence
    2. Lower the model temperature for all requests
    3. Log all user messages for later review
    4. Instruct the model to refuse requests for its instructions and add an output filter that detects verbatim system-prompt content Correct answer.

    Explanation

    The vulnerability is disclosure of confidential instructions, so the fix combines a refusal instruction with detection of leaked prompt text in outputs. Temperature, prompt length, and logging do not prevent the disclosure itself.

  165. Question 165Test, monitor, and optimize AI solutions

    When comparing two models on a benchmark, a team reports only the mean score. What is the most important addition for interpreting whether the difference is meaningful?

    1. The number of parameters in each model
    2. The wall-clock time of the evaluation run
    3. The file size of each model's weights
    4. Confidence intervals or standard error computed across evaluation samples Correct answer.

    Explanation

    A mean alone hides sampling variability, so confidence intervals show whether an observed gap could be noise. Weight size, run duration, and parameter count describe the models or the process but not the reliability of the comparison.

  166. Question 166Test, monitor, and optimize AI solutions

    Which practice best limits the risk that an evaluation set becomes contaminated by content the model was trained on?

    1. Encrypt the evaluation set at rest
    2. Store the evaluation set in a different cloud region
    3. Run the evaluation more frequently
    4. Track provenance of evaluation items and check for overlap with training corpora Correct answer.

    Explanation

    Contamination occurs when evaluation items appear in training data, inflating scores. Recording provenance and checking overlap detects this. Storage location, encryption, and run frequency do not address overlap with training content.

  167. Question 167Test, monitor, and optimize AI solutions

    A team wants each production request to be traceable from user input through retrieval, prompt assembly, and model response for later debugging. Which practice is most appropriate?

    1. Record only aggregate daily request counts
    2. Emit structured trace records with a shared request identifier across each pipeline stage Correct answer.
    3. Capture only failed requests
    4. Log only the final model response text

    Explanation

    A shared request identifier lets engineers reconstruct the exact path a request took, including which documents were retrieved and how the prompt was built. Logging only responses or aggregates loses the intermediate context, and failed-only capture misses successful-but-wrong cases.

  168. Question 168Test, monitor, and optimize AI solutions

    A safety classifier that filters model outputs flags 30 percent of benign responses while catching nearly all harmful ones. Which adjustment best improves the user experience without abandoning safety goals?

    1. Lower the threshold further to catch more harmful outputs
    2. Apply the classifier only to responses longer than 500 tokens
    3. Remove the classifier entirely to eliminate false positives
    4. Raise the decision threshold and re-measure both false-positive and false-negative rates on a labeled set Correct answer.

    Explanation

    The classifier is over-blocking, so raising the threshold trades some recall for precision. Re-measuring both error types on labeled data confirms the tradeoff is acceptable. Removing the filter abandons safety, lowering the threshold worsens over-blocking, and length-based filtering is unrelated to risk.

  169. Question 169Test, monitor, and optimize AI solutions

    An evaluation of a multi-step agent reports a high task success rate, but logs show many runs succeed only after several failed tool calls. Which additional metric best captures this inefficiency?

    1. Average number of tool calls and retries per successful task Correct answer.
    2. Number of distinct tools available to the agent
    3. Average size of the tool schema definitions
    4. Average length of the final answer text

    Explanation

    Success rate hides the path taken, while tool-call and retry counts reveal wasted steps and cost. Answer length, tool inventory, and schema size describe outputs or configuration rather than execution efficiency.

  170. Question 170Test, monitor, and optimize AI solutions

    Which metric most directly measures whether a toxicity filter applied to generated text is working as intended?

    1. Average response latency added by the filter
    2. Number of filters deployed per environment
    3. Percentage of requests that reached the model
    4. Rate of flagged outputs confirmed as toxic by human review Correct answer.

    Explanation

    Confirming flagged outputs against human judgment measures whether the filter's decisions are correct. Latency, deployment counts, and request routing describe operations rather than filter accuracy.

  171. Question 171Secure, govern, and deploy AI solutions

    A team must ensure that only approved model deployments can be created in an Azure AI Foundry project. Which governance control enforces this requirement?

    1. Azure Policy assignments scoped to the resource group Correct answer.
    2. Role-based access control at the subscription level
    3. Resource locks of type CanNotDelete
    4. Diagnostic settings forwarded to a Log Analytics workspace

    Explanation

    Azure Policy evaluates resource properties during creation and can deny deployments that violate allowed values, which is how approved-model-only rules are enforced. Resource locks and RBAC do not validate model properties, and diagnostic settings only record activity after the fact.

  172. Question 172Secure, govern, and deploy AI solutions

    You need to give a data scientist permission to read and query an Azure AI Search index used by a RAG solution, following least privilege. Which built-in role should you assign?

    1. Search Index Data Reader Correct answer.
    2. Search Service Contributor
    3. Owner
    4. Cognitive Services OpenAI User

    Explanation

    Search Index Data Reader grants query-only access to index data, which matches the least-privilege requirement. Contributor and Owner roles manage the service rather than granting data-plane query access, and the OpenAI role does not apply to Azure AI Search.

  173. Question 173Secure, govern, and deploy AI solutions

    A RAG application must store document embeddings in a way that supports vector similarity search with metadata filters. Which Azure service is purpose-built for this?

    1. Azure AI Search Correct answer.
    2. Azure Event Hubs
    3. Azure Blob Storage
    4. Azure Cosmos DB for Table

    Explanation

    Azure AI Search provides native vector and hybrid search with filterable metadata fields, making it the purpose-built choice for retrieval in RAG. Blob Storage stores objects but does not perform similarity search, and the other services are not search engines.

  174. Question 174Secure, govern, and deploy AI solutions

    A security reviewer asks how to detect when someone changes the content safety filter configuration of an Azure AI resource. Which capability provides this audit trail?

    1. Azure Monitor activity log Correct answer.
    2. Azure Advisor recommendations
    3. Service Health notifications
    4. Application Insights sampling

    Explanation

    The activity log records control-plane operations such as configuration changes to Azure resources, providing the audit trail. Advisor and Service Health are advisory, and Application Insights sampling affects telemetry volume rather than recording management operations.

  175. Question 175Secure, govern, and deploy AI solutions

    An AI solution must ensure that prompt and completion content is never used to improve the underlying foundation model. Which configuration satisfies this?

    1. Increase the tokens-per-minute quota for the deployment
    2. Disable the abuse-monitoring data retention option where available and document the data-handling policy Correct answer.
    3. Enable content filtering at the highest severity threshold
    4. Turn on diagnostic logging to a storage account

    Explanation

    Opting out of data retention for abuse monitoring, where the service offers it, prevents prompt and completion content from being stored or reused for model improvement. Content filtering, logging, and quota changes do not affect whether data is retained for training or monitoring.

  176. Question 176Secure, govern, and deploy AI solutions

    You deploy a model to Azure AI Foundry and need to restrict outbound network access so the endpoint is reachable only from a specified virtual network. Which feature should you configure?

    1. CORS rules on the endpoint
    2. A shared access signature on the storage account
    3. Private endpoint with public network access disabled Correct answer.
    4. API key rotation policy

    Explanation

    A private endpoint places the resource on a virtual network, and disabling public network access ensures the endpoint is not reachable from the internet. SAS tokens, CORS, and key rotation affect access credentials or browser behavior, not network reachability.

  177. Question 177Secure, govern, and deploy AI solutions

    A compliance officer requires that every AI response be traceable to the specific documents retrieved for it. Which design element best satisfies this?

    1. Raise the temperature setting for the model
    2. Return citations that reference the source document identifiers and chunks used in grounding Correct answer.
    3. Store the raw prompt in a local browser cache
    4. Enable autoscale on the model deployment

    Explanation

    Citations that identify the retrieved chunks give a verifiable link between an answer and its grounding sources, which is the basis for traceability. Caching, temperature, and autoscale affect performance or output variability, not provenance.

  178. Question 178Secure, govern, and deploy AI solutions

    Your organization wants to compare the performance of two prompt variants before promoting one to production. Which practice supports responsible, governed deployment?

    1. Promote the variant with the longest responses
    2. Skip evaluation because prompt changes are low risk
    3. Evaluate both variants against a labeled test set and review metrics before promotion Correct answer.
    4. Deploy both variants directly to production and let users choose

    Explanation

    Offline evaluation against a labeled dataset produces comparable, reviewable evidence that supports a governed promotion decision. Shipping untested variants or using response length as a proxy does not measure quality or safety.

  179. Question 179Secure, govern, and deploy AI solutions

    You must prevent an AI application from returning content that violates a category policy while still allowing benign discussion of sensitive topics. Which approach is most appropriate?

    1. Block all requests containing any sensitive keyword
    2. Route all traffic through a manual review queue
    3. Configure content safety filters with severity thresholds tuned per category Correct answer.
    4. Disable content filtering entirely

    Explanation

    Per-category severity thresholds let you block harmful content while permitting legitimate discussion, balancing safety with usability. Keyword blocking is overbroad, disabling filters removes protection, and manual review of all traffic does not scale.

  180. Question 180Secure, govern, and deploy AI solutions

    A team needs to log model inputs and outputs for auditing while minimizing exposure of personal data. Which technique should be applied before logging?

    1. Redaction or pseudonymization of detected sensitive entities Correct answer.
    2. Base64 encoding of the payload
    3. Compression of the log records
    4. Increasing log retention to ten years

    Explanation

    Redaction or pseudonymization removes or masks personal data so logs remain useful for audit without exposing sensitive information. Encoding and compression are reversible transformations that do not protect the data, and longer retention increases exposure.

  181. Question 181Secure, govern, and deploy AI solutions

    You are designing a multi-tenant AI application where each tenant's data must remain isolated. Which control most directly enforces logical data isolation at query time?

    1. Tenant-scoped filters applied to every retrieval query Correct answer.
    2. One API key shared by all tenants
    3. A global cache keyed only by prompt text
    4. A single shared index for all tenants

    Explanation

    Applying tenant-scoped filters to every retrieval query ensures results are constrained to the requesting tenant's data. Shared indexes without filters, shared keys, and prompt-only caches can leak data across tenants.

  182. Question 182Secure, govern, and deploy AI solutions

    An auditor asks how you can prove that a deployed model has not been silently swapped. Which control provides the strongest evidence?

    1. Enable autoscaling on the deployment
    2. Rotate the endpoint key monthly
    3. Rely on the model name only
    4. Pin the model version in the deployment and record the version in release metadata Correct answer.

    Explanation

    Pinning an explicit model version and recording it in release metadata creates verifiable evidence of exactly what was deployed. Model names can point to moving targets, and autoscale or key rotation do not identify the model version.

  183. Question 183Secure, govern, and deploy AI solutions

    Your organization requires that AI-generated content be identifiable in downstream systems. Which practice supports this?

    1. Increase the max tokens setting
    2. Attach provenance metadata or a watermark to generated outputs Correct answer.
    3. Store outputs only in memory
    4. Lower the frequency penalty

    Explanation

    Provenance metadata or watermarking lets downstream systems recognize content as AI-generated, supporting transparency obligations. Token limits and penalties affect generation style, and in-memory storage prevents downstream identification.

  184. Question 184Secure, govern, and deploy AI solutions

    A developer needs to call an Azure AI service from a CI/CD pipeline without storing long-lived secrets. Which approach aligns with governance best practice?

    1. Use a managed identity with federated credentials for the pipeline Correct answer.
    2. Store the key in a public repository variable
    3. Embed the API key in the pipeline YAML
    4. Email the key to the pipeline owner

    Explanation

    Managed identities with workload identity federation let pipelines obtain short-lived tokens without persisted secrets. Embedding keys in YAML or repositories exposes them, and emailing keys is insecure.

  185. Question 185Secure, govern, and deploy AI solutions

    You must ensure that an AI solution's data stays within a specific geography for regulatory reasons. Which deployment decision addresses this?

    1. Enable geo-redundant replication to all regions
    2. Use a global endpoint for lowest latency
    3. Turn on cross-region inferencing
    4. Deploy the AI resource and its dependent storage in an approved region Correct answer.

    Explanation

    Placing the AI resource and its dependent data stores in an approved region keeps data within the required geography. Geo-replication, global endpoints, and cross-region inferencing can move data outside the approved boundary.

  186. Question 186Secure, govern, and deploy AI solutions

    A team wants to catch regressions in answer quality after a model upgrade. Which monitoring approach is most appropriate?

    1. Disable logging to reduce noise
    2. Rely on user complaints alone
    3. Run automated evaluations against a golden dataset on a schedule and alert on metric drops Correct answer.
    4. Monitor only CPU utilization

    Explanation

    Scheduled evaluation against a stable golden dataset detects quality regressions with objective metrics and enables alerting. Complaints are lagging and sparse, infrastructure metrics do not measure quality, and disabling logging removes evidence.

  187. Question 187Secure, govern, and deploy AI solutions

    Which Azure capability helps identify potentially harmful or jailbreak-style prompts before they reach the model?

    1. Prompt shields in Azure AI Content Safety Correct answer.
    2. Virtual network peering
    3. Azure Advisor cost recommendations
    4. Blob lifecycle management

    Explanation

    Prompt shields detect prompt-injection and jailbreak attempts so they can be blocked before inference. Advisor, lifecycle policies, and peering address cost, storage, and networking rather than prompt risk.

  188. Question 188Secure, govern, and deploy AI solutions

    You need to grant an application permission to call an Azure OpenAI model without granting the ability to manage the resource. Which role assignment meets this?

    1. User Access Administrator
    2. Cognitive Services Contributor
    3. Cognitive Services OpenAI User Correct answer.
    4. Subscription Owner

    Explanation

    Cognitive Services OpenAI User grants data-plane inference access without management rights. Contributor and Owner roles include management permissions, and User Access Administrator manages role assignments.

  189. Question 189Secure, govern, and deploy AI solutions

    A RAG solution must ensure that retrieved documents are current and that outdated documents are excluded. Which design choice best supports this?

    1. Increase the model temperature
    2. Store effective and expiration dates as filterable metadata and filter at query time Correct answer.
    3. Rely on the model's training cutoff
    4. Retrieve a fixed number of documents regardless of date

    Explanation

    Date metadata with query-time filters lets the retriever exclude expired documents deterministically. Temperature and fixed retrieval counts do not address currency, and training cutoffs are unrelated to stored documents.

  190. Question 190Secure, govern, and deploy AI solutions

    An organization wants to review AI system behavior and decisions with stakeholders on a recurring basis. Which artifact best supports this governance activity?

    1. A maintained system card documenting intended use, limitations, and evaluation results Correct answer.
    2. A list of API keys
    3. A screenshot of the deployment page
    4. An unstructured chat transcript

    Explanation

    A system card records intended use, known limitations, and evaluation evidence, giving reviewers a structured basis for governance decisions. Screenshots and transcripts are incomplete, and key lists are a security liability.

  191. Question 191Secure, govern, and deploy AI solutions

    You must limit the impact of a compromised API key used by an AI application. Which two-part approach is most effective?

    1. Share one key across all environments and rotate annually
    2. Disable logging to avoid exposing the key
    3. Store the key in source control for easy access
    4. Use short-lived credentials from a managed identity and enforce least-privilege roles Correct answer.

    Explanation

    Short-lived credentials reduce the window of misuse, and least-privilege roles limit what a compromised identity can do. Shared long-lived keys, source-control storage, and disabled logging all increase risk.

  192. Question 192Secure, govern, and deploy AI solutions

    A model deployment must handle traffic spikes without manual intervention while capping cost. Which configuration addresses both goals?

    1. Remove all quota limits
    2. Use the lowest possible capacity at all times
    3. Set a fixed high capacity for all hours
    4. Configure autoscale with a defined maximum capacity Correct answer.

    Explanation

    Autoscale adjusts capacity to demand, and a maximum cap bounds cost during spikes. Fixed high capacity wastes spend, removing quotas risks runaway cost, and minimal capacity fails under load.

  193. Question 193Secure, govern, and deploy AI solutions

    A regulated workload requires that all data in transit to the AI endpoint be encrypted and that clients be authenticated. Which configuration satisfies both?

    1. Enable CORS for all origins
    2. Enforce TLS 1.2 or later and require Microsoft Entra ID token authentication Correct answer.
    3. Allow HTTP and use a shared API key
    4. Disable authentication for internal networks

    Explanation

    Enforcing modern TLS protects data in transit, and Entra ID token authentication verifies client identity. Plain HTTP, shared keys alone, permissive CORS, and disabled authentication weaken one or both requirements.

  194. Question 194Secure, govern, and deploy AI solutions

    Your team must ensure that evaluation datasets containing sensitive records can be used for testing without exposing raw values to evaluators. Which approach is best?

    1. Share the raw production dataset with all evaluators
    2. Use de-identified or synthetic datasets representative of production distributions Correct answer.
    3. Post the dataset in a public repository
    4. Skip evaluation to avoid handling sensitive data

    Explanation

    De-identified or synthetic data preserves representative test coverage while preventing exposure of sensitive values. Sharing raw data, publishing it, or skipping evaluation creates privacy or quality risk.

  195. Question 195Secure, govern, and deploy AI solutions

    An AI agent is allowed to call tools that modify records. Which control best limits unintended actions?

    1. Allow all tools without logging
    2. Grant the agent administrator rights to avoid failures
    3. Cache tool results to skip approvals
    4. Require human approval for high-impact tool calls and scope tool permissions to the minimum needed Correct answer.

    Explanation

    Human approval for high-impact actions plus least-privilege tool scopes constrains what the agent can do if it misbehaves. Broad rights, unlogged tools, and cached approvals all increase the risk of unintended changes.

  196. Question 196Secure, govern, and deploy AI solutions

    A team wants to confirm that a generative AI application's responses meet a factual-accuracy bar before it is released to customers. Which approach provides the most direct evidence for that release decision?

    1. Configure autoscaling on the deployment so that evaluation traffic is handled without throttling.
    2. Run an offline evaluation against a curated ground-truth dataset and compare the aggregate quality scores to an agreed threshold. Correct answer.
    3. Enable diagnostic logging on the model deployment and review the request volume over the past week.
    4. Increase the model's temperature setting so responses read more naturally to reviewers.

    Explanation

    Release gating for quality requires a measurement against known-good reference data, which offline evaluation with ground-truth labels provides. Comparing aggregate metrics to a pre-agreed threshold turns the measurement into a defensible go/no-go decision. Logging volume, sampling parameters, and scale settings do not measure answer correctness.

  197. Question 197Secure, govern, and deploy AI solutions

    Which Azure AI Foundry capability is designed to intercept and block harmful or adversarial prompts before they are passed to the underlying model?

    1. Model benchmark leaderboards
    2. Deployment token rate limits
    3. Content filters configured on the model deployment Correct answer.
    4. Prompt flow batch runs

    Explanation

    Content filtering evaluates prompts and completions against harm categories and can block a request before it reaches the model. The other listed features govern orchestration, comparison, or throughput rather than content safety.

  198. Question 198Secure, govern, and deploy AI solutions

    An AI application must call an Azure AI service from automated pipelines. Security policy forbids storing client secrets in pipeline variables. Which configuration satisfies the policy?

    1. A personal access token issued to an administrator and reused across pipelines
    2. A service principal federated to the pipeline's identity provider, granted the least-privileged role on the resource Correct answer.
    3. A shared account password rotated every ninety days and stored in a pipeline variable group
    4. A long-lived API key written into the repository's encrypted secret file

    Explanation

    Workload identity federation lets the pipeline exchange its own platform-issued identity token for an access token, so no client secret exists to store or leak. Pairing this with a least-privileged role assignment limits what the pipeline can do. Every other option depends on a durable credential that must be stored somewhere.

  199. Question 199Secure, govern, and deploy AI solutions

    A regulated workload requires that every request to an AI endpoint be encrypted in transit and that each caller be individually authenticated. Which combination meets both requirements?

    1. A private endpoint plus a shared subscription key distributed to all callers
    2. TLS enforcement on the endpoint plus per-caller Microsoft Entra ID authentication Correct answer.
    3. Azure Monitor alerts plus diagnostic logging of request metadata
    4. Network security group rules plus IP allowlisting for the client range

    Explanation

    Transport encryption is satisfied by enforcing TLS on the endpoint, and individual caller authentication is satisfied by requiring each caller to present its own Entra ID token. Together these address both stated requirements. Network controls, shared keys, and monitoring each address only part of the problem.

  200. Question 200Secure, govern, and deploy AI solutions

    An AI agent may invoke tools that update records in a business system. Which control most directly limits the blast radius of an unintended or manipulated tool call?

    1. Increasing the agent's context window so it can reason over more of the conversation history
    2. Enabling verbose logging of every tool invocation for later review by the operations team
    3. Lowering the model temperature so the agent's tool selections become more deterministic
    4. Requiring human approval for tool calls that write or delete records, while allowing read-only calls to proceed Correct answer.

    Explanation

    A human-in-the-loop gate on state-changing operations prevents an erroneous or injected instruction from altering data without review, which is the most direct way to cap impact. Read-only calls can stay automated because they cannot change records. The other options influence behavior or visibility but do not prevent a harmful write.

AB-620 flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

Deterministic logic placement
Rule stating that fixed, rule-based calculations and validations belong in application code rather than in a model prompt, because code is testable and produces identical output every time.
Latency budget
The maximum acceptable end-to-end response time, decomposed across model inference, retrieval, and network overhead so each component has a measurable allocation.
Context window
The total number of tokens a model can process in one request, covering system prompt, retrieved content, conversation history, and the generated output combined.
RAG versus fine-tuning
Decision rule: retrieval suits changing or citable facts, while fine-tuning suits stable knowledge where the gap is tone, format, or task-specific behavior.
Chunk overlap
A repeated portion of adjacent text chunks, typically ten to twenty percent, that prevents answers from being lost when relevant content straddles a chunk boundary.
Human-in-the-loop trigger
A condition that routes a model output to a person for approval, such as low confidence, an irreversible action, or a high-risk intent classification.
Tool schema
The typed definition of a function exposed to an agent, including parameter names, types, descriptions, and required flags that the model uses to form valid calls.
Agent iteration cap
An explicit maximum on plan-act-observe cycles in an agent loop, preventing runaway token spend and infinite retry behavior when tools keep failing.
Workflow versus agent
Design choice: use a fixed workflow when steps are known in advance, and an autonomous agent only when the next step depends on runtime results.
Grounded response
An answer whose factual claims are supported by a retrieved record or tool response, with the supporting source identified rather than asserted from model memory.
Idempotency key
A caller-supplied identifier on a state-changing tool call that lets the service detect retries and avoid creating duplicate records or repeating side effects.
Structured handoff summary
A compact object passed between agents containing intent, constraints, and collected facts, replacing the raw transcript to control token growth and preserve focus.
Fallback chain
An ordered sequence of alternatives, such as primary model then secondary model then scripted response, each hop logging its reason for later analysis.
Retrieval pipeline order
The fixed sequence of ingest, chunk, embed, index, query embedding, similarity search, optional rerank, and prompt assembly used in retrieval-augmented generation.
Hybrid search
Combining vector similarity with keyword matching and merging ranked results, which improves recall for queries containing exact identifiers, codes, or product names.
Embedding model consistency
The requirement that query and document vectors come from the same model and version, meaning the corpus must be re-embedded whenever the embedding model changes.
Exponential backoff with jitter
A retry strategy that increases wait time between attempts and adds randomization, reducing synchronized retry storms when a service is throttling requests.
Chunk metadata
Fields attached to each chunk such as source id, title, section path, timestamp, and access tags, enabling citation, freshness checks, and per-user filtering.
Event-driven index update
Indexing triggered by source-system change events rather than a nightly batch, which shortens the staleness window between source updates and retrievable content.
Data contract validation
Checking incoming payloads against a defined schema, nullability, and value ranges before sending them to a model, rejecting payloads that fail validation.
Groundedness metric
An evaluation score measuring the proportion of claims in a response that are supported by the provided context or retrieved sources.
Retrieval recall
The fraction of queries whose relevant chunk appears within the top-k retrieved results, a leading indicator of downstream answer quality.
Regression gate
A release rule requiring that a prompt or model change not push any per-category evaluation metric below its defined minimum threshold.
Drift detection
Monitoring input topic and output length distributions against a baseline, triggering re-evaluation when observed shifts exceed expected variation.
Optimization order
A prioritization sequence for quality work: retrieval quality first, then prompt clarity, then model selection, then caching, because retrieval errors dominate failures.
Prompt caching TTL
A time-to-live on cached model responses tied to data freshness, disabled for personalized or permission-filtered answers that must not be reused across users.
Single-variable A/B test
An experiment changing exactly one factor, such as prompt, model, temperature, or retrieval depth, so any measured difference can be attributed to that factor.
Prompt injection defense
Treating retrieved documents and user input as untrusted data rather than instructions, with tool permissions enforced outside the model in code.
Least privilege agent identity
A design in which each agent holds only the scopes its tools require, so a compromised or manipulated model cannot reach data beyond the workflow's needs.
PII redaction
Detecting and removing sensitive personal fields before sending content to a model, and logging identifiers or hashes instead of raw values.
Canary release
Deploying a new model or prompt version to a small slice of live traffic first, monitoring quality and error signals before full promotion.
Audit logging
Recording prompt version, model version, tool calls, and decision outcome for each request so any output can be reconstructed and reviewed later.
Layered content safety
Applying input filtering, model-side safety settings, and output filtering together, since no single control reliably blocks all harmful content.

AB-620 study guide

Plan and Design AI Solutions

Start every design from the business outcome, not the model. A defensible plan names the user, the decision or task being assisted, the success metric, and the failure cost. Distinguish deterministic automation (rules, lookups) from generative or predictive AI; many requirements that look like AI are better served by deterministic logic, and exam items often reward the simpler design.

Know the core design vocabulary: use-case fit (classification, extraction, summarization, retrieval-augmented generation, agentic task execution), data readiness (availability, quality, labeling, freshness, licensing), and non-functional constraints (latency, throughput, cost per request, availability, explainability, residency). Model selection is a trade-off exercise: capability vs. latency vs. cost vs. context window. Prefer the smallest model that meets the quality bar, and plan an evaluation set before committing.

Traps to watch for:

  • Choosing a complex agentic design when a single retrieval-grounded prompt satisfies the requirement.
  • Ignoring data governance (PII, retention, licensing) until after architecture is fixed.
  • Treating a prototype's demo quality as evidence of production readiness.
  • Forgetting human-in-the-loop for high-cost or regulated decisions.

Rule of thumb: if you cannot state the metric that proves the solution works and the guardrail that stops it from causing harm, the design is not finished. Worked example: a request to "auto-approve expense reports" should decompose into extraction, policy retrieval, a confidence threshold, and human review for low-confidence or high-value cases - not a single free-form prompt that decides payment.

Document assumptions, constraints, and rejected alternatives; exam scenarios frequently test whether you can justify a design choice against stated constraints.

Build and Configure AI Agents and Workflows

An agent is a model plus instructions, tools, memory, and a control loop. Configuration quality usually matters more than model choice. Write system instructions that define role, scope, tone, refusal behavior, and output format. Keep instructions specific and testable; vague persona text is a common cause of unreliable behavior.

Understand tool/function calling: each tool needs a clear name, description, typed parameters, and error contract. The model selects tools from descriptions, so ambiguous descriptions cause wrong calls. Design for idempotency and validate tool inputs before execution. Orchestration patterns include single-call, chained prompts, routing to specialized prompts, and multi-step agent loops with planning and reflection. More autonomy means more latency, cost, and unpredictability - bound loops with max-iteration limits and termination conditions.

Memory and context management: conversation history, retrieved documents, and summarized state all consume the context window. Use summarization or structured state to keep long workflows coherent. Grounding via retrieval reduces hallucination; require citations or structured output where accuracy matters.

Traps:

  • Instructions that conflict with tool descriptions or output schema.
  • Unbounded agent loops with no stop condition or budget.
  • Assuming the model will infer missing parameter constraints.
  • Mixing planning and execution without a validation step between them.

Rule of thumb: every agent step should be independently testable with a fixed input and an asserted output. Worked example: for a support triage agent, configure a router prompt that classifies intent, a retrieval tool scoped to the knowledge base, and a handoff tool for billing disputes - each with explicit schemas and a fallback path when confidence is low.

Integrate AI Solutions with Data and Services

Integration is where AI solutions meet reality: data sources, APIs, identity, and failure modes. For retrieval-augmented generation, understand the pipeline - ingestion, chunking, embedding, indexing, retrieval, reranking, prompt assembly. Chunk size and overlap affect recall and precision; metadata filters scope retrieval to the right tenant or document class. Evaluate retrieval separately from generation, because a correct answer from wrong context is still a defect.

For service integration, know authentication patterns (managed identity or scoped keys over embedded secrets), rate limiting with retry and exponential backoff, timeouts, and circuit breaking. Design for partial failure: a downstream API outage should degrade gracefully, not crash the workflow. Normalize and validate data at boundaries; treat model output as untrusted input to downstream systems.

Structured output (JSON schemas, typed responses) makes model output machine-consumable and testable. Prefer schema-constrained output over parsing free text.

Traps:

  • Storing credentials in prompts or client code.
  • Assuming retrieval returns the right document without metadata scoping (cross-tenant leakage).
  • Retrying non-idempotent writes and duplicating side effects.
  • Passing raw model text into a database or API without validation.

Rule of thumb: validate at every boundary - data in, model out, and service responses. Worked example: a document Q&A service should filter the index by tenant and access level before retrieval, rerank the top candidates, require the model to cite chunk IDs, and reject answers whose citations fall outside the retrieved set.

Test, Monitor, and Optimize AI Solutions

AI systems fail silently, so testing must cover both deterministic behavior and probabilistic quality. Build a golden evaluation set of representative inputs with expected outputs or grading criteria. Use automated metrics where objective (exact match, schema validity, retrieval recall) and model-based or human grading for subjective qualities (helpfulness, faithfulness, tone). Track quality per category, not just an aggregate average that hides regressions.

Regression testing runs the eval set on every prompt, model, or retrieval change. Version prompts and configurations alongside code so changes are attributable. Include adversarial and edge cases: empty input, very long input, prompt injection attempts, out-of-scope requests, and ambiguous queries.

Production monitoring covers latency, error rates, token usage and cost, tool-call failures, and quality signals from user feedback or sampled review. Log prompts and responses with appropriate redaction and retention. Define alert thresholds on both system health and quality drift.

Optimization levers: prompt refinement, retrieval tuning, caching, smaller models for easy routes, batching, and streaming for perceived latency. Optimize against the metric you defined, not vibes.

Traps:

  • Reporting only an average score and missing a category regression.
  • Changing the model without rerunning evals.
  • Logging sensitive data without redaction.
  • Treating a passing eval set as proof of general correctness.

Rule of thumb: no change ships without an eval run and a recorded baseline. Worked example: after switching models, rerun the golden set; if faithfulness improves but latency doubles, route simple queries to the smaller model and keep the larger one for complex cases.

Secure, Govern, and Deploy AI Solutions

Security and governance are design requirements, not a final checklist. Start with identity and least privilege: the AI service should access only the data and tools it needs, using managed identities and scoped roles. Separate the model's permissions from the user's - enforce authorization at the data layer, not in the prompt. Apply defense in depth against prompt injection: treat retrieved content and user input as untrusted, constrain tool permissions, and require confirmation for high-impact actions.

Data protection covers classification, PII detection and redaction, encryption in transit and at rest, retention limits, and residency. Know where data flows: prompts, logs, retrieval indexes, and third-party model endpoints. Governance includes an inventory of AI use cases, risk assessment, human oversight for consequential decisions, audit trails, and clear accountability.

Deployment patterns: managed endpoints vs. self-hosted, blue/green or canary rollout with rollback, environment separation (dev/test/prod with distinct credentials), and infrastructure-as-code for reproducibility. Use canary releases to limit blast radius when prompts or models change.

Traps:

  • Relying on prompt instructions as a security control.
  • Logging full prompts containing sensitive data.
  • Granting the agent broad write access for convenience.
  • Deploying prompt changes without canary or rollback.

Rule of thumb: assume the model can be manipulated; enforce every critical control outside the model. Worked example: for an agent that issues refunds, cap refund amount in the tool implementation, require human approval above a threshold, log every action with actor identity, and scope credentials to the refund API only.

AB-620 cheat sheet

Quick reference

Plan and Design AI Solutions

  • Requirement triage rule: classify each stated need as deterministic logic, classification, extraction, or generative - deterministic logic belongs in code, not in a model prompt.

  • Latency budget formula: end-to-end p95 = model p95 + retrieval p95 + network overhead; a design fails the SLA if any single component's p95 exceeds the total budget.

  • Context window sizing: tokens needed = system prompt + retrieved chunks + conversation history + max output; if the sum exceeds the window, reduce chunk count before reducing output limit.

  • Model selection order: capability threshold first, then latency, then cost per 1K tokens, then regional availability; a cheaper model below the capability threshold is not a valid option.

  • RAG vs fine-tuning rule: choose retrieval when facts change frequently or must be cited; choose fine-tuning when the gap is tone, format, or task-specific behavior on stable knowledge.

  • Chunking defaults: 200-800 tokens per chunk with 10-20% overlap; overlap exists to prevent answer loss at chunk boundaries, not to increase recall of unrelated content.

  • Cost estimate formula: monthly cost = (input tokens x input price) + (output tokens x output price), summed over expected request volume, including retries and evaluation traffic.

  • Human-in-the-loop trigger design: require review when model confidence is below threshold, when the action is irreversible, or when the request matches a high-risk intent class.

Quick reference

Build and Configure AI Agents and Workflows

  • Tool schema rule: every tool parameter needs a type, a description, and a required flag; ambiguous descriptions are the most common cause of malformed tool calls.

  • Agent loop pattern: observe, plan, call tool, evaluate result, repeat - cap iterations explicitly because unbounded loops are the primary driver of runaway token spend.

  • Orchestration choice: use a fixed workflow when steps are known in advance; use an autonomous agent only when the next step depends on runtime results.

  • Grounding rule: an agent must cite the retrieved record or tool response that supports a factual claim; ungrounded claims are treated as hallucinations regardless of plausibility.

  • Handoff design: pass a structured summary object between agents (intent, constraints, collected facts) rather than the raw transcript, to control token growth.

  • Idempotency requirement: any tool that creates or mutates state must accept an idempotency key so retries after a timeout do not duplicate side effects.

  • Fallback chain: primary model, then secondary model, then deterministic scripted response; each hop must log the reason for the fallback for later tuning.

  • Prompt structure order: role and constraints, then task, then context, then output format; output format instructions placed last are followed more reliably.

Quick reference

Integrate AI Solutions with Data and Services

  • Retrieval pipeline stages in order: ingest, chunk, embed, index, query-embed, similarity search, optional rerank, then prompt assembly.

  • Hybrid search rule: combine vector similarity with keyword search and merge with reciprocal rank fusion when queries contain exact identifiers, codes, or product names.

  • Embedding consistency requirement: query and document vectors must come from the same model and version; re-embed the corpus whenever the embedding model changes.

  • API integration pattern: wrap model calls in a service layer that handles auth, retries with exponential backoff, timeouts, and response validation before downstream use.

  • Freshness design: index updates should be event-driven from the source system; a nightly batch alone creates a staleness window equal to the batch interval.

  • Chunk metadata minimum: source id, document title, section path, last-modified timestamp, and access-control tags; access tags are required to filter results per user.

  • Rate limit handling: read the retry-after header when present, otherwise back off exponentially with jitter; fixed-interval retries amplify throttling.

  • Data contract rule: define the schema, nullability, and expected value ranges for every field passed to the model, and reject payloads that fail validation.

Quick reference

Test, Monitor, and Optimize AI Solutions

  • Evaluation set design: hold out a labeled set covering common, edge, and adversarial cases; report per-category scores because an aggregate score hides regressions.

  • Core metrics: groundedness (claims supported by context), relevance (answer addresses the question), and retrieval recall (relevant chunk present in top-k).

  • Regression gate rule: a prompt or model change ships only if it does not decrease any per-category metric below its defined floor.

  • Online monitoring signals: latency percentiles, error and refusal rate, token usage per request, tool-call failure rate, and user thumbs-down rate.

  • Drift detection: track input topic distribution and output length distribution over time; a shift beyond the baseline band triggers re-evaluation.

  • Optimization order: fix retrieval quality first, then prompt clarity, then model choice, then caching - retrieval errors dominate most quality failures.

  • Caching rule: cache identical normalized prompts with a TTL tied to data freshness; disable caching for personalized or permission-filtered responses.

  • A/B comparison rule: change one variable at a time (prompt, model, temperature, retrieval depth) so the observed delta is attributable.

Quick reference

Secure, Govern, and Deploy AI Solutions

  • Prompt injection defense: treat all retrieved and user-supplied content as untrusted data, never as instructions; enforce tool permissions outside the model.

  • Least privilege rule: each agent identity gets only the scopes its tools require; the model must not hold credentials that allow broader access than the workflow needs.

  • PII handling: detect and redact sensitive fields before sending to the model, and log only identifiers or hashes rather than raw content.

  • Deployment gate: require evaluation pass, security review, and a rollback plan before promotion; canary release to a small traffic slice first.

  • Audit requirement: log prompt version, model version, tool calls, and decision outcome for every request so any output can be reconstructed.

  • Content safety layering: apply input filtering, model-side safety settings, and output filtering; no single layer is sufficient on its own.

  • Data residency rule: route requests and store logs only in regions permitted by policy; verify the deployment's region configuration matches the data classification.

  • Responsible AI disclosure: users must be told they are interacting with an AI system, and high-impact decisions must retain a documented human review path.

AB-620 key terms

Retrieval-Augmented Generation (RAG)

A pattern where a model's answer is grounded by first retrieving relevant documents or records from a data source and inserting them into the prompt, reducing hallucination and keeping responses current without retraining.

Related: Grounding, Vector Index, Chunking
Grounding

Anchoring model output to verifiable source content, such as retrieved documents or system records, so answers cite or reflect real data rather than relying only on pretrained knowledge.

Related: Retrieval-Augmented Generation (RAG), Citation
Chunking

Splitting large documents into smaller passages before embedding and indexing, balancing chunk size so retrieved context is specific enough to be relevant yet large enough to be meaningful.

Related: Embedding, Vector Index, Retrieval-Augmented Generation (RAG)
Embedding

A numeric vector representation of text or other content that captures semantic meaning, letting systems compare items by similarity for search, clustering, or recommendation.

Related: Vector Index, Semantic Search, Chunking
Vector Index

A data structure that stores embeddings and supports fast approximate nearest-neighbor lookup, enabling semantic retrieval over large document collections at query time.

Related: Embedding, Semantic Search, Retrieval-Augmented Generation (RAG)
Semantic Search

Retrieval based on meaning similarity between a query and stored content rather than exact keyword matches, typically powered by embeddings and a vector index.

Related: Embedding, Vector Index, Hybrid Search
Hybrid Search

A retrieval approach that combines keyword matching with vector similarity and merges the ranked results, improving recall for exact terms while preserving semantic relevance.

Related: Semantic Search, Vector Index
System Prompt

Instructions configured ahead of user input that define an agent's role, tone, constraints, and available tools, shaping behavior across every conversation turn.

Related: Agent, Tool Calling, Guardrail
Agent

An AI component that receives a goal, reasons about next steps, invokes tools or models, and iterates until it produces a result or reaches a stopping condition.

Related: Tool Calling, Orchestration, System Prompt
Tool Calling

A mechanism where a model requests that an external function, API, or service be executed with specific arguments, then uses the returned result to continue its reasoning.

Related: Agent, Function Schema, Orchestration
Function Schema

A structured description of a callable tool, including its name, purpose, and typed parameters, that tells the model when and how to invoke it correctly.

Related: Tool Calling, Agent
Orchestration

The coordination of multiple steps, models, or agents in a defined sequence or branching flow, managing state, retries, and handoffs between stages.

Related: Agent, Workflow, Human-in-the-Loop
Workflow

A predefined sequence of automated steps, triggers, and conditions that processes inputs and produces outputs, often combining model calls with deterministic business logic.

Related: Orchestration, Trigger, Idempotency
Human-in-the-Loop

A design where a person reviews, approves, or corrects an AI output at a defined checkpoint before the process continues, used for high-risk or low-confidence actions.

Related: Orchestration, Guardrail, Confidence Threshold
Prompt Template

A reusable prompt with placeholders that are filled at runtime with user input, retrieved context, or variables, keeping prompts consistent and maintainable across calls.

Related: System Prompt, Few-Shot Prompting
Few-Shot Prompting

Including a small number of worked examples in a prompt so the model can infer the desired format, tone, or reasoning pattern for the current task.

Related: Prompt Template, System Prompt
Connector

A configured integration that lets an AI solution read from or write to an external system such as a database, file store, or SaaS application using that system's protocol and credentials.

Related: API Endpoint, Managed Identity, Data Source
API Endpoint

A specific network address and method through which a service exposes operations, used by agents or workflows to send requests and receive structured responses.

Related: Connector, Rate Limiting, Function Schema
Rate Limiting

A control that caps how many requests a client may send in a time window, protecting downstream services and keeping integrations within provider quotas.

Related: API Endpoint, Throttling, Retry Policy
Retry Policy

Rules governing how failed calls are reattempted, including maximum attempts, delay between tries, and which error types qualify, often using exponential backoff.

Related: Rate Limiting, Idempotency, Throttling
Idempotency

A property where repeating the same request produces the same effect as sending it once, preventing duplicate records or actions when retries occur.

Related: Retry Policy, Workflow
Data Source

The origin system or store that supplies content to an AI solution, such as a database, document library, or message queue, along with its access rules.

Related: Connector, Indexing Pipeline
Indexing Pipeline

An automated process that ingests source content, extracts and chunks text, generates embeddings, and writes them to a search index on a schedule or on change.

Related: Chunking, Embedding, Data Source
Evaluation Dataset

A curated set of representative inputs paired with expected outputs or grading criteria, used to measure solution quality and compare changes objectively.

Related: Regression Testing, Groundedness Metric, Golden Response
Groundedness Metric

A score measuring how well a generated answer is supported by the provided source content, flagging responses that assert facts absent from the evidence.

Related: Evaluation Dataset, Hallucination, Citation
Regression Testing

Re-running an evaluation dataset after a change to prompts, models, or data to confirm quality has not degraded and previously passing cases still succeed.

Related: Evaluation Dataset, Golden Response
Golden Response

A human-reviewed reference answer treated as the expected result for a test case, used as a baseline when scoring model output.

Related: Evaluation Dataset, Regression Testing
Hallucination

Confidently stated output that is factually wrong or unsupported by available evidence, often caused by missing context or overgeneralization from training data.

Related: Groundedness Metric, Grounding, Citation
Token

A unit of text, roughly a word fragment, that models process and bill against; token counts determine context limits and cost per request.

Related: Context Window, Latency
Context Window

The maximum number of tokens a model can consider in a single request, spanning system instructions, conversation history, retrieved content, and the response.

Related: Token, Chunking
Latency

The elapsed time between sending a request and receiving a usable response, influenced by model choice, output length, retrieval steps, and network overhead.

Related: Token, Caching
Caching

Storing previously computed responses or retrieved results so repeated or similar requests can be served faster and at lower cost.

Related: Latency, Rate Limiting
Confidence Threshold

A cutoff on a model or classifier score that determines whether an output is accepted automatically, escalated for review, or rejected.

Related: Human-in-the-Loop, Guardrail
Guardrail

A configured control that constrains AI behavior, such as blocking harmful content, restricting topics, or validating outputs before they reach users or downstream systems.

Related: Content Filter, System Prompt, Human-in-the-Loop
Content Filter

A screening layer that detects and blocks or annotates harmful, unsafe, or disallowed material in prompts and responses according to policy categories and severity levels.

Related: Guardrail, Responsible AI
Responsible AI

A set of principles and practices for building AI that is fair, transparent, accountable, and safe, guiding design decisions about harms, disclosure, and oversight.

Related: Content Filter, Guardrail, Data Residency
Managed Identity

A platform-issued identity assigned to a workload that obtains credentials automatically, letting the solution access protected resources without storing secrets in code.

Related: Role-Based Access Control (RBAC), Key Vault, Connector
Role-Based Access Control (RBAC)

An authorization model that grants permissions based on assigned roles rather than individual accounts, simplifying least-privilege management across users and services.

Related: Managed Identity, Key Vault, Data Residency
Key Vault

A secured service for storing and auditing secrets, keys, and certificates so applications retrieve them at runtime instead of embedding credentials in configuration files.

Related: Managed Identity, Role-Based Access Control (RBAC)
Data Residency

The requirement that data be stored and processed only in approved geographic regions, affecting where models, indexes, and logs may be deployed.

Related: Responsible AI, Role-Based Access Control (RBAC)

AB-620 study plan

8 weeks · About 7 hours per week

Week 1

Build and configure AI agents and workflows

  • Map the full lifecycle of an agent (instructions, tools, triggers, channels) and diagram it

  • Build a minimal working agent that answers from a single connected source

  • Define at least three tool/action capabilities and wire one into the agent

  • Log what configuration choices changed agent behavior and why

Your checkpoint

A working single-source agent plus a one-page architecture diagram and behavior-change log

Week 2

Build and configure AI agents and workflows & Plan and design AI solutions

  • Design a multi-step workflow with branching logic and at least one human-approval checkpoint

  • Translate three business requirements into agent configuration decisions

  • Document where design constraints (latency, cost, data access) force trade-offs

Your checkpoint

A workflow spec with branch diagram and a requirements-to-configuration traceability table

Week 3

Plan and design AI solutions

  • Compare three solution patterns (single agent, orchestrated workflow, retrieval-augmented) against a stated scenario

  • Write design decisions that justify source selection and data-flow boundaries

  • Identify failure modes for each pattern and propose a mitigation for each

Your checkpoint

A design decision record covering pattern choice, risks, and mitigations for one scenario

Week 4

Integrate AI solutions with data and services

  • Connect the agent to two distinct data sources and normalize their outputs

  • Implement a retrieval step that returns grounded citations

  • Handle an API error and a timeout with explicit fallback behavior

  • Document connector authentication and permission requirements

Your checkpoint

An integrated solution reading from two sources with citations and demonstrated error fallback

Week 5

Integrate AI solutions with data and services & Test, monitor, and optimize AI solutions

  • Build a 15-case evaluation set spanning happy path, edge cases, and adversarial inputs

  • Run the evaluation and record pass/fail with reasons

  • Instrument logging for inputs, tool calls, and outputs

  • Tune at least one configuration parameter and measure the change

Your checkpoint

An evaluation report with baseline scores, a tuning change, and post-change scores

Week 6

Test, monitor, and optimize AI solutions

  • Define three monitoring signals (quality, latency, cost) and set alert thresholds

  • Create a regression test that catches a previously observed failure

  • Analyze one cost driver and reduce it without dropping below quality bar

Your checkpoint

A monitoring dashboard outline plus a passing regression suite and a cost-reduction note

Week 7

Secure, govern, and deploy AI solutions

  • Apply least-privilege access to agent identity and connected data

  • Classify data handled by the solution and document handling rules per class

  • Configure content safety and prompt-injection guards, then test them

  • Draft an audit trail requirement list

Your checkpoint

A security and governance checklist completed against the built solution, with guard test results

Week 8

Secure, govern, and deploy AI solutions & Test, monitor, and optimize AI solutions & Plan and design AI solutions

  • Package the solution for deployment with environment separation

  • Run a full end-to-end validation in a pre-production environment

  • Re-run the evaluation set and confirm no regression

  • Write a go-live and rollback runbook

Your checkpoint

A deployed pre-production solution with final evaluation scores and a go-live/rollback runbook

Milestones

Week 2: Working agent and a multi-step workflow both function end to end without manual intervention

Week 4: Solution retrieves grounded answers from two sources and survives an injected API failure

Week 6: Evaluation suite passes at an agreed threshold and monitoring signals are defined with thresholds

Week 8: Pre-production validation passes with no regression and a complete runbook exists

AB-620 exam day

Your checklist

Government-issued photo ID

Bring the exact form of ID the testing provider requires; verify name spelling matches your registration. Confirm acceptable ID types with the provider before test day.

Appointment confirmation

Have your registration confirmation, appointment time, and test-center location (or online-proctoring access link) saved offline on your phone and printed as backup.

Login credentials for remote testing

If testing online, run the provider's system check in advance and confirm your username/password and any required software are working. Confirm proctoring requirements with the provider.

Approved calculator or on-screen equivalent

Check the exam's official policy on calculators and reference materials; do not assume a physical calculator is allowed. Confirm with the provider.

Quiet, well-lit test space

For remote testing, clear your desk of notes and devices and ensure a stable internet connection. Review the provider's room and environment rules beforehand.

Review the published exam guide and content outline

Re-read the official exam objectives and scoring/format notes so you know question types, time limit, and whether breaks are allowed. Confirm specifics with the provider.

Plan your route and arrival buffer

For in-person testing, arrive with time to spare for check-in and ID verification; account for traffic and parking.

Rest and hydration plan

Protect sleep the night before and plan a light meal so you are not distracted by hunger or fatigue during the test.

Backup contact for the test center

Save the provider's support phone number in case of technical issues or a scheduling problem.

Plan the day

Final 72 hours

Do a final pass over weak topics using practice questions; stop learning brand-new material. Confirm your appointment details, ID requirements, and (for remote testing) run the system check. Re-read the official exam guide for format and rules.

Final 24 hours

Light review of summary notes and key formulas only. Prepare your ID, confirmation, and any approved materials in one place. Confirm the start time and location/link. Get a full night's sleep; avoid cramming late.

Morning of

Eat a normal meal, arrive early (or log in early for remote testing), and do a short warm-up of a few easy practice questions to settle nerves. Recheck that your desk and environment meet the rules.

During the test

Read each question fully, watch the clock, and flag uncertain items to revisit. Answer every question if there is no penalty for guessing (confirm the scoring policy with the provider). Use allowed breaks strategically and stay calm if a technical issue arises-note it and contact support.

Rules to check

  • Bring valid government-issued photo ID; the name must match your registration. Confirm accepted ID forms with the provider.

  • Follow all proctor instructions and the provider's rules on personal items, notes, phones, and workspace; unauthorized materials can lead to score cancellation.

  • Do not communicate with others or use unauthorized aids during the exam; treat all content as confidential per the provider's agreement.

  • Confirm the break policy and whether breaks count against your time before test day; rules vary by delivery method.

  • Understand the scoring model in advance, including whether unanswered questions are penalized, by checking the official exam guide.

  • If a technical problem occurs during remote testing, stay on camera if required and contact support promptly; document what happened.

  • Arrive or log in early enough to complete check-in and identity verification without eating into your test time.

AB-620 FAQ

How many AB-620 practice questions are available?

The practice set has 200 AB-620 questions. Each one includes the choices, the correct answer, and an explanation, and the set is free to read with no signup.

How many AB-620 flashcards are available?

The flashcard set has 33 AB-620 cards. Each card shows the term and its definition.

Which topics does the AB-620 study material cover?

The AB-620 study material has 5 sections, including Plan and Design AI Solutions, Build and Configure AI Agents and Workflows, and Integrate AI Solutions with Data and Services.