I85PREP
⌘ K
AI-103 study materials
AI-103 study resources

AI-103 practice questions

One question at a time. Build your AI-103 understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Plan and manage an Azure AI solution

A company wants to call several Azure AI services from one application and manage a single set of credentials. Which resource type should the company create?

  1. A container instance for each AI service
  2. A separate single-service resource for each AI service
  3. A multi-service Azure AI services resource Correct answer.
  4. An Azure Machine Learning workspace

Explanation

A multi-service Azure AI services resource exposes one endpoint and one key set that can be used across multiple AI services, which simplifies credential management. Separate single-service resources would each need their own endpoint and keys.

Question 2Plan and manage an Azure AI solution

An administrator must retrieve the endpoint and access keys for an Azure AI services resource. Where in the Azure portal can the administrator find them?

  1. The Networking blade of the resource
  2. The Identity blade of the resource
  3. The Keys and Endpoint page Correct answer.
  4. The Export template blade of the resource

Explanation

The Keys and Endpoint page displays the resource endpoint and the two subscription keys used for key-based authentication. Identity and Networking blades control managed identities and network access respectively.

Question 3Plan and manage an Azure AI solution

Which authentication approach does Microsoft recommend for applications calling Azure AI services in production?

  1. Subscription keys passed in the request header
  2. Shared access signatures generated per request
  3. Microsoft Entra ID with a managed identity Correct answer.
  4. Connection strings embedded in application code

Explanation

Microsoft Entra ID, ideally combined with a managed identity, issues short-lived tokens and removes the need to distribute long-lived secrets. Keys and connection strings are secrets that must be protected and rotated.

Question 4Plan and manage an Azure AI solution

What is the primary purpose of an Azure resource group?

  1. To control which users can sign in to the portal
  2. To replicate resources across multiple regions
  3. To define the network boundary for private endpoints
  4. To group resources that share a lifecycle Correct answer.

Explanation

A resource group is a logical container for resources that share a lifecycle, and deleting the group deletes its members. It is neither a network boundary nor an identity boundary.

Question 5Plan and manage an Azure AI solution

Which two factors are most directly affected by the region selected for an Azure AI services resource?

  1. The network latency and the data residency Correct answer.
  2. The role assignments and the resource tags
  3. The number of API keys issued and the SLA credit
  4. The pricing tier and the model version

Explanation

Region selection determines where data is stored and processed, which affects data residency commitments, and how far traffic travels, which affects latency. Pricing tier and keys are configured separately.

Question 6Plan and manage an Azure AI solution

A team deploys an Azure AI services resource on the F0 pricing tier. What limitation should the team expect?

  1. The resource cannot use Microsoft Entra ID authentication.
  2. It is capped at a monthly transaction quota. Correct answer.
  3. The resource cannot be monitored with Azure Monitor.
  4. The resource must be deployed in two regions.

Explanation

The F0 free tier is intended for evaluation and enforces a monthly transaction limit in addition to rate limits. It still supports Entra ID authentication, monitoring, and deployment to a single region.

Question 7Plan and manage an Azure AI solution

A finance team wants to attribute Azure AI costs to individual departments. Which feature should be used?

  1. Resource tags that are included in cost analysis Correct answer.
  2. Role assignments scoped to each department
  3. Diagnostic settings that forward metrics
  4. Azure Policy definitions applied at the subscription

Explanation

Tags are key-value pairs that flow into cost analysis and invoices, so charges can be grouped by department. They do not grant or restrict access.

Question 8Plan and manage an Azure AI solution

An engineer wants to retain request logs from an Azure AI services resource for later analysis. Which destination can diagnostic settings send logs to?

  1. An Azure App Configuration store
  2. An Azure Log Analytics workspace Correct answer.
  3. An Azure Container Registry instance
  4. An Azure Key Vault instance

Explanation

Diagnostic settings route platform logs and metrics to a Log Analytics workspace, a storage account, or an event hub. A workspace supports querying and retention for later analysis.

Question 9Plan and manage an Azure AI solution

An App Service web app must call an Azure AI services resource without storing any secret. What should be configured?

  1. A service principal secret rotated every 30 days
  2. A shared access signature stored in app settings
  3. A managed identity with a role assignment Correct answer.
  4. The resource's primary key in Azure Key Vault

Explanation

A managed identity gives the app an Entra ID identity, and a role assignment on the AI resource authorizes calls without any stored secret. Retrieving a key from Key Vault still relies on a secret.

Question 10Plan and manage an Azure AI solution

A company must ensure that an Azure AI services resource is reachable only from a specific virtual network. Which configuration should be applied?

  1. A service endpoint on the virtual network subnet
  2. A private endpoint with public access disabled Correct answer.
  3. A firewall rule that allows the corporate IP range
  4. A network security group on the resource subnet

Explanation

A private endpoint places a private IP for the service in the subnet, and disabling public network access blocks internet traffic. Together they restrict access to the specified virtual network.

Question 11Plan and manage an Azure AI solution

What is required before an Azure AI services resource can use customer-managed keys for encryption at rest?

  1. A Key Vault key that the resource is authorized to use Correct answer.
  2. A private endpoint connected to Azure Monitor
  3. A storage account in the same subscription as the resource
  4. A managed identity assigned to the resource group

Explanation

Customer-managed keys require a Key Vault or Managed HSM that stores the key, plus a managed identity on the AI resource with permission to use that key. The key wraps the service-managed data encryption key.

Question 12Plan and manage an Azure AI solution

Which requirement applies to running Azure AI containers in a disconnected environment?

  1. The container must be deployed to Azure Kubernetes Service.
  2. The container must use the free F0 pricing tier.
  3. It requires a commitment plan and Microsoft approval. Correct answer.
  4. The container must reach the billing endpoint every 24 hours.

Explanation

Disconnected containers run without internet access and require approval for disconnected use plus billing under a commitment plan. They do not contact the billing endpoint.

Question 13Plan and manage an Azure AI solution

When running an Azure AI services container on-premises, which two values must be provided to the container for billing?

  1. The resource group name and the region
  2. The billing endpoint URI and an API key Correct answer.
  3. The model name and the deployment name
  4. The subscription ID and the tenant ID

Explanation

Containers need the billing endpoint URI and an API key so usage can be reported to the Azure resource. The other values are not container runtime billing inputs.

Question 14Plan and manage an Azure AI solution

An application is receiving 429 responses from an Azure AI services resource during peak load. What is the most direct first step to resolve the throttling?

  1. Raise the pricing tier or request more quota Correct answer.
  2. Move the resource to a different resource group
  3. Rotate the resource's secondary key
  4. Enable diagnostic logging on the resource

Explanation

429 responses indicate rate limiting, and raising the tier or quota increases the allowed request rate. Key rotation and logging do not change throughput limits.

Question 15Plan and manage an Azure AI solution

An organization wants to prevent creation of Azure AI services resources outside approved regions. Which Azure capability should be used?

  1. An Azure Policy with a location restriction Correct answer.
  2. Azure Advisor recommendations
  3. A network security group rule
  4. A resource lock of type ReadOnly

Explanation

Azure Policy can deny deployments that do not match allowed locations, enforcing regional restrictions at scale. Locks and network rules do not evaluate resource location.

Question 16Plan and manage an Azure AI solution

A team wants to be alerted before Azure AI spending for a subscription exceeds a defined amount. Which feature should they configure?

  1. An action group attached to a metric alert
  2. A resource lock on the subscription
  3. A diagnostic setting for the resource
  4. A budget with an alert threshold Correct answer.

Explanation

Budgets in Cost Management trigger alerts at defined thresholds, warning teams before spending exceeds the target. Metric alerts track resource telemetry rather than spend.

Question 17Plan and manage an Azure AI solution

A team deploys Azure AI resources with an Azure Resource Manager template. What does redeploying the same template with the same parameters do?

  1. It deletes resources that are not in the template by default.
  2. It creates a second copy of each resource in a new region.
  3. It fails because resource names must be unique per deployment.
  4. It reaches the same resource state without duplicates. Correct answer.

Explanation

ARM deployments are idempotent, so redeploying the same template with the same parameters produces the same state instead of duplicate resources. Only resources explicitly defined are managed.

Question 18Plan and manage an Azure AI solution

A developer needs to call an Azure AI services endpoint but must not be able to read or regenerate the resource keys. Which role assignment meets this requirement?

  1. Cognitive Services User on the resource Correct answer.
  2. User Access Administrator on the subscription
  3. Contributor on the resource
  4. Owner on the resource group

Explanation

The Cognitive Services User role allows calls to the service with Entra ID credentials but does not include key management. Owner and Contributor can list and regenerate keys.

Question 19Plan and manage an Azure AI solution

What is the key distinction between a virtual network service endpoint and a private endpoint for Azure AI services?

  1. Only a private endpoint requires a resource group in the same region.
  2. Only a service endpoint supports Microsoft Entra ID authentication.
  3. A private endpoint assigns a private IP; a service endpoint does not. Correct answer.
  4. A service endpoint encrypts traffic; a private endpoint does not.

Explanation

A private endpoint gives the service a private IP address inside the subnet, which is the basis for fully private access. A service endpoint keeps traffic on the Azure backbone, but the service still has a public IP address.

Question 20Plan and manage an Azure AI solution

An operations team wants to track the latency of calls to an Azure AI services resource over time. Which capability provides this data?

  1. Resource locks in the portal
  2. Azure Policy compliance reports
  3. Metrics in Azure Monitor Correct answer.
  4. The Export template feature

Explanation

Azure Monitor metrics store numeric time-series data such as latency and error counts, which can be charted and alerted on. Policy, locks, and template export do not provide telemetry.

Question 21Plan and manage an Azure AI solution

A workload requires predictable, low-latency inference and a fixed monthly cost for an Azure OpenAI model. Which deployment option best fits?

  1. Serverless API deployment billed per token
  2. A batch deployment that processes queued files
  3. Provisioned throughput deployment Correct answer.
  4. Standard deployment with pay-as-you-go billing

Explanation

Provisioned throughput reserves model capacity and is billed at a fixed rate, giving predictable latency for steady high-volume workloads. Standard deployments are pay-as-you-go and can be throttled.

Question 22Plan and manage an Azure AI solution

An application must survive a datacenter failure in its primary region while keeping all data in that same geography. Which design choice should be evaluated first?

  1. Creating a second resource in an unrelated geography
  2. Deploying with availability zone support in a zonal region Correct answer.
  3. Relying on the service's automatic key failover
  4. Using the free tier in two subscriptions

Explanation

Availability zones spread replicas across physically separate datacenters within a region, protecting against a datacenter failure while keeping data in the same geography. Not every region or service supports zones.

Question 23Plan and manage an Azure AI solution

An administrator must rotate keys for an Azure AI services resource with zero downtime. What is the recommended sequence?

  1. Disable local authentication, then re-enable it with new keys.
  2. Delete the resource and redeploy it with new keys.
  3. Move clients to key2, regenerate key1, then move clients back. Correct answer.
  4. Regenerate the primary key, then immediately regenerate the secondary key.

Explanation

Azure AI services issue two keys so clients can run on one key while the other is regenerated. Moving clients to the secondary key first avoids downtime.

Question 24Plan and manage an Azure AI solution

In Azure AI Foundry, what is the relationship between a hub and a project?

  1. A hub shares connections and resources with many projects. Correct answer.
  2. A hub is a compute cluster that a single project uses.
  3. A hub and a project are the same resource with different names.
  4. A project can contain multiple hubs from different subscriptions.

Explanation

A hub is a shared container for connections, compute, and resources, and projects created under it inherit those shared assets. Multiple projects can be created in one hub.

Question 25Plan and manage an Azure AI solution

A security team requires that all calls to an Azure AI services resource use Microsoft Entra ID and that key-based authentication is blocked. Which configuration enforces this?

  1. Configure a private endpoint for the resource.
  2. Enable diagnostic logging for authentication events.
  3. Remove the Contributor role from all developers.
  4. Disable local authentication on the resource. Correct answer.

Explanation

The resource property that disables local authentication forces all callers to use Microsoft Entra ID and rejects key-based requests. Role changes and logging do not block key use.

Question 26Plan and manage an Azure AI solution

An Azure AI services resource in production must be protected from accidental deletion by any user, including subscription owners. Which Azure capability should be used?

  1. A diagnostic setting that forwards activity logs to a workspace
  2. A resource lock with the CanNotDelete level Correct answer.
  3. A role assignment that removes the Contributor role from all users
  4. An Azure Policy definition that audits resources missing required tags

Explanation

Resource locks are management-plane settings applied directly to a resource that block delete operations regardless of the caller's RBAC permissions, and the lock must be removed before deletion can succeed. Azure Policy governs whether resources may be created or changed, and logging only records what happened.

Question 27Plan and manage an Azure AI solution

Before an application can send prompts to a model hosted in an Azure AI Foundry resource, what must exist in that resource?

  1. A model deployment Correct answer.
  2. A vector index in the project
  3. A private endpoint attached to the project
  4. An evaluation run for the model

Explanation

Applications call a deployment endpoint rather than a model in the catalog, so a deployment must be created first; it fixes the model, version, capacity, and hosting type. The other artifacts are optional features that are not on the request path.

Question 28Plan and manage an Azure AI solution

An application consistently reaches the default tokens-per-minute quota on its Azure OpenAI deployment. What is the supported way to obtain more capacity?

  1. Enable diagnostic logging on the resource
  2. Add a second key to the Azure AI services resource
  3. Change the deployment to the Batch processing type
  4. Submit a quota increase request for the region Correct answer.

Explanation

Quota is allocated per subscription, region, and model, and increases are granted through the quota request process before the limit changes. Keys, deployment type, and logging settings do not raise the throughput ceiling for real-time calls.

Question 29Plan and manage an Azure AI solution

An administrator wants to know whether a platform-level incident in the resource's region is affecting an Azure AI services resource. Which Azure capability provides this information?

  1. Azure Service Health Correct answer.
  2. Metric alerts configured on the resource
  3. Availability tests in Application Insights
  4. The activity log of the resource group

Explanation

Service Health reports Microsoft-side incidents, planned maintenance, and advisories scoped to the subscriptions, services, and regions you care about. The other capabilities report your own telemetry and control-plane operations rather than Azure platform status.

Question 30Plan and manage an Azure AI solution

A solution needs a single credential and a single base endpoint for speech, language, and vision operations, with all usage billed together. Which resource should be created?

  1. A separate resource for each of the three services
  2. A multi-service Azure AI services resource Correct answer.
  3. An Azure Machine Learning workspace
  4. An Azure OpenAI resource in each required region

Explanation

A multi-service Azure AI services resource exposes one endpoint and one key across multiple prebuilt services and consolidates their billing, which matches the stated requirement. Per-service resources needed separate endpoints and keys, and the other resource kinds do not provide these prebuilt services together.

Question 31Plan and manage an Azure AI solution

Several Azure AI services resources in different resource groups must share one identity whose permissions are granted before any of those resources is created. Which identity type should be used?

  1. The resource's primary access key shared between resources
  2. A user-assigned managed identity Correct answer.
  3. A system-assigned managed identity enabled on each resource
  4. A service principal whose client secret is stored in each resource's configuration

Explanation

A user-assigned managed identity is a standalone Azure resource that can be created and granted roles ahead of time, then attached to any number of AI resources. System-assigned identities live and die with a single resource, so roles cannot be granted in advance and each resource gets a different identity.

Question 32Plan and manage an Azure AI solution

A company has 40 subscriptions and wants every new AI resource to inherit the same role assignments and policy assignments without repeating configuration per subscription. What should be used?

  1. A resource group in each subscription
  2. A policy definition assigned per subscription
  3. An Azure Advisor recommendation applied to each resource
  4. A management group containing the subscriptions Correct answer.

Explanation

Management groups provide a scope above subscriptions, and role and policy assignments made there are inherited by every subscription and resource beneath them. Assigning the same configuration separately per subscription or resource group repeats work and can drift.

Question 33Plan and manage an Azure AI solution

A workload sends a steady, high volume of Speech-to-text requests every month and needs a lower unit price than pay-as-you-go billing. Which option should be selected?

  1. Provisioned throughput units on an Azure OpenAI deployment
  2. The free F0 pricing tier
  3. A commitment tier for the Azure AI services resource Correct answer.
  4. A reserved instance on the underlying virtual machines

Explanation

Commitment tiers are prepaid monthly usage levels for Azure AI services that reduce the per-unit price as sustained volume grows. Reserve-based and provisioned options apply to other resource types and do not discount Speech transactions.

Question 34Plan and manage an Azure AI solution

A team must process a very large document set asynchronously overnight and wants the lowest cost per token. Which deployment type should be used?

  1. Provisioned deployment with reserved throughput
  2. Global Batch deployment Correct answer.
  3. A Docker container running the model on-premises
  4. Standard deployment with a high tokens-per-minute quota

Explanation

Batch deployments accept large asynchronous jobs on a separate, discounted quota and write results to a storage location, which suits non-interactive overnight processing. Real-time and provisioned deployments are priced and sized for latency-sensitive traffic.

Question 35Plan and manage an Azure AI solution

A regulated customer requires that prompts and completions for an Azure OpenAI model be processed only in the region where the deployment was created. Which deployment type should be used?

  1. Global Standard
  2. Standard Correct answer.
  3. Global Batch
  4. Data Zone Standard

Explanation

A Standard deployment processes requests in the region that hosts the deployment, which is the only option here that guarantees regional processing. Global Standard and Global Batch may route to any region where the model is available, and Data Zone Standard confines processing to the wider data zone.

Question 36Plan and manage an Azure AI solution

A healthcare customer requires that no prompts or completions be stored or reviewed by Microsoft for abuse monitoring. What must the customer do?

  1. Configure a customer-managed key for the resource
  2. Request and receive an exemption from abuse monitoring Correct answer.
  3. Set the content filter severity threshold to the highest level
  4. Disable diagnostic settings on the resource

Explanation

Abuse monitoring retention and human review can be turned off only through an approved exemption granted to eligible customers, because it is a Microsoft-operated control rather than a resource setting. Content filters, encryption keys, and logging settings affect different concerns and do not change abuse monitoring.

Question 37Plan and manage an Azure AI solution

A project in Azure AI Foundry must use an existing Azure AI Search index without hardcoding its endpoint and credentials. Which artifact should be created in the project?

  1. A model deployment that points to the search service
  2. A managed private endpoint to the search service
  3. A connection to the Azure AI Search resource Correct answer.
  4. A vector index created in the project's storage

Explanation

A connection stores the target endpoint and the authentication method for an external resource so project tools can use it without embedding secrets in application code. Deployments host models, and private endpoints only affect network reachability.

Question 38Plan and manage an Azure AI solution

An application calls the Azure AI services REST API and must not be affected by changes to preview features. What should each request specify?

  1. An Ocp-Apim-Subscription-Region header for the deployment
  2. The newest preview api-version so features stay current
  3. The model version string in place of the API version
  4. A generally available api-version query parameter Correct answer.

Explanation

The api-version query parameter selects the API contract that governs request and response shapes, so pinning a generally available version isolates the application from preview changes. The model version controls model behavior, and the region header only identifies where the resource lives.

Question 39Plan and manage an Azure AI solution

A team has unused tokens-per-minute quota for a model in West Europe but keeps hitting its limit for the same model in East US. What should the team do?

  1. Raise the East US deployment capacity above the regional quota
  2. Add a secondary key to the resource to increase throughput
  3. Open a support request to relocate the East US quota
  4. Deploy the model in West Europe and route some traffic there Correct answer.

Explanation

Quota is allocated per subscription, region, and model, so capacity that is unused in one region can only be consumed by deploying the model in that region and directing traffic to it. Quota cannot be transferred between regions or exceeded through keys or deployment settings.

Question 40Plan and manage an Azure AI solution

An application depends on an Azure OpenAI model version that Microsoft has announced will be retired. What is the recommended planning action?

  1. Switch the deployment to the F0 tier to extend the retirement date
  2. Enable diagnostic logging to keep the version available
  3. Nothing, because retired model versions keep serving existing deployments
  4. Migrate to a supported model version before the retirement date Correct answer.

Explanation

Model versions follow a published retirement schedule, and requests to a retired version fail after that date, so the application must be tested and moved to a supported version beforehand. Pricing tier and logging settings have no influence on the schedule.

Question 41Implement generative AI and agentic solutions

In a chat completions request, which message role is used to set the assistant's persona, tone, and behavioral constraints?

  1. assistant
  2. tool
  3. user
  4. system Correct answer.

Explanation

The system message is processed as high-level instruction that shapes how the model responds throughout the conversation. User messages carry the human input, assistant messages carry prior model output, and tool messages carry results returned from function calls.

Question 42Implement generative AI and agentic solutions

What is the main purpose of creating embeddings for documents in a retrieval-augmented generation solution?

  1. It compresses each document so that it consumes less storage space in the index
  2. It translates each document into the language of the end user
  3. It maps text to vectors that place semantically similar text close together Correct answer.
  4. It encrypts each document so that only the model can read it

Explanation

Embeddings are numeric vectors produced by an embedding model; text with similar meaning produces vectors that are near each other under a distance metric. Vector similarity search then retrieves passages whose meaning matches the user's question rather than only matching literal keywords.

Question 43Implement generative AI and agentic solutions

Where do you configure the maximum number of tokens per minute that a model deployment can process?

  1. In the system message of every chat request that is sent
  2. In the deployment's tokens-per-minute rate limit Correct answer.
  3. In the temperature value that is set on each request
  4. In the content filter severity threshold

Explanation

Rate limits are a property of the deployment, expressed as tokens per minute, and requests that exceed that limit are throttled. Prompt-level settings such as the system message and temperature affect content and style but never change capacity.

Question 44Implement generative AI and agentic solutions

Which HTTP status code indicates that a request was rejected because the deployment's rate limit was exceeded?

  1. 401
  2. 200
  3. 400
  4. 429 Correct answer.

Explanation

Status 429 means too many requests and signals throttling, so the client should retry after the interval indicated by the Retry-After header. Handling 429 with exponential backoff is the standard resilience pattern for generative AI workloads.

Question 45Implement generative AI and agentic solutions

When a model returns a function call in a chat completions response, which component executes the underlying function?

  1. The model runtime executes the function and injects the result into the next response automatically
  2. The content filter executes the function after validating its parameters
  3. The deployment's rate limiter executes the function during the next throttling window
  4. The client application runs the function and returns the result Correct answer.

Explanation

Function calling is a contract: the model only emits the function name and JSON arguments, and the calling application performs the actual work. The application then submits the function output back to the model so it can produce the final natural-language answer.

Question 46Implement generative AI and agentic solutions

Which change most directly reduces fabricated answers in a retrieval-augmented generation application?

  1. Shorten the system message to a single sentence
  2. Increase the temperature so the model explores more candidate answers
  3. Enable a presence penalty on the completion
  4. Include the retrieved documents in the prompt as grounding context Correct answer.

Explanation

Grounding the prompt with retrieved source passages gives the model factual material to cite and constrains it from inventing details. Temperature and penalties change sampling behavior but do not supply missing facts.

Question 47Implement generative AI and agentic solutions

A knowledge base contains documents that are longer than the model's context window. What is the recommended way to make them searchable?

  1. Truncate every document to its first few hundred tokens before indexing
  2. Deploy the largest available model so that the context window grows
  3. Split the documents into overlapping chunks and index each chunk separately Correct answer.
  4. Raise the temperature until the model can generalize across the document

Explanation

Chunking divides long documents into passages that fit the embedding model and the prompt budget, and overlapping the boundaries preserves continuity between adjacent chunks. Each chunk is embedded and indexed independently so retrieval can return only the passages that matter.

Question 48Implement generative AI and agentic solutions

Which harm categories are graded for severity by the default Azure OpenAI content filters?

  1. Hate, sexual, violence, and self-harm Correct answer.
  2. Sentiment, intent, topic, and language
  3. PII, PHI, financial data, and credentials
  4. Prompt, completion, embedding, and image

Explanation

The default content filtering system classifies text and images into hate, sexual, violence, and self-harm categories, each with severity levels that map to block or annotate behavior. Separate features such as prompt shields and PII detection address injection attacks and personal data.

Question 49Implement generative AI and agentic solutions

A RAG application must match exact product codes and also handle users who describe products in their own words. Which Azure AI Search configuration is most appropriate?

  1. Hybrid search combining keyword and vector queries with semantic ranking enabled Correct answer.
  2. Pure vector search with a very large top-k and no keyword matching at all
  3. Keyword search only, with fuzziness enabled on every searchable field
  4. A custom scoring profile that boosts the most recently indexed documents

Explanation

Hybrid search runs keyword and vector queries together and fuses the results, so literal identifiers are matched exactly while paraphrased descriptions are matched semantically. Semantic ranking then reranks the fused result set using a language model to improve top results.

Question 50Implement generative AI and agentic solutions

In the Foundry Agent Service, which object stores the sequence of messages exchanged between a user and an agent?

  1. A vector store
  2. A thread Correct answer.
  3. A tool
  4. A run

Explanation

A thread is the durable conversation record that accumulates user and assistant messages, and each processing step against that thread is a run. The agent reads the thread's messages, invokes tools as needed, and appends its response to the same thread.