You have a Microsoft Foundry project that serves a high-volume chat app.Most requests are simple FAQs, but some require advanced reasoning.You need to reduce costs and latency for common queries, without degrading the quality of the responses to complex questions.What should you do?
Answer(s): B
One should absolutely use a model cascade to route requests to different models based on complexity. This architectural pattern is highly effective for high-volume chat applications because it directly addresses the trade-off between operational cost, API latency, and response quality.Using a model cascade router ensures that your high-volume Microsoft Foundry application scales efficiently by reserving expensive computational power exclusively for queries that actually require advanced cognitive processing.Note: 1. Analyze Request Complexity Implement a lightweight intent classifier or routing layer at the entry point of your Microsoft Foundry project. This router quickly inspects incoming user prompts using basic heuristic keyword matching, semantic embeddings, or a highly optimized, fast model (like Phi-3 or GPT-4o-mini) to categorize the query as either a "Simple FAQ" or a "Complex Reasoning" request.2. Route to the Optimal TierTier 1 (Fast & Cheap): Route standard, predictable FAQ requests to a smaller, cost-effective model or a local cache/vector database lookup. This keeps latency in milliseconds and drastically lowers token costs.Tier 2 (Advanced Reasoning): Route multi-step logic, coding, or highly contextual queries to a frontier model (like GPT-4o).3. Implement Fallback LogicDesign the cascade to be dynamic. If the smaller Tier 1 model generates a response with low confidence, or if the user asks a follow-up question that invalidates the simple FAQ status, seamlessly upgrade the conversation loop to the Tier 2 model.
https://medium.com/@sujathamudadla1213/what-is-the-primary-purpose-of-a-model-cascade-in-machine-learning-0b145a7bc6e2
HOTSPOT (Drag and Drop is not supported)You have a Microsoft Foundry project that contains an internal Q&A agent.Users report the following issues when they ask the agent questions:-An increase in the following response: “No relevant information found”-Periodic HTTP 429 rate limit exceeded errors during peak hoursYou need to identify whether each issue is caused by model unavailability, resource limits, or inference failures.What should you do? To answer, select the appropriate options in the answer area.Note: Each correct selection is worth one point.Hot Area:
Answer(s): A
Box 1: Model Availability rate and Provisioned Utilization. Metrics to enableTo diagnose these issues, the Model Availability rate and Provisioned Utilization metrics are needed.Problem 1 ("No relevant information found"): This points to an inference failure (RAG hallucination or retrieval gap), which is often caused by the model hallucinating when it is missing contextual grounding.Problem 2 (HTTP 429 rate limit exceeded): This indicates resource limits. Monitoring provisioned utilization helps identify if your allocated capacity is simply maxing out during peak hours.Box 2: RequestResponse Diagnostic log to collectTo isolate whether your issues are caused by model unavailability, resource limits, or inference failures, you need to collect and analyze the RequestResponse diagnostic log category (also referred to as RequestResponseLogs).While your enabled metrics provide high-level context (e.g., Model Availability highlights backend service drops and Provisioned Utilization highlights capacity strain), they do not reveal per-request details. The RequestResponse log records the exact HTTP status codes, error codes, and backend latency for every transaction.
https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/monitor-models
You have a Microsoft Foundry project that contains a high-traffic agent.After a recent update, operational costs increase significantly.Monitoring confirms that the volume of user traffic to the agent remains unchanged.You suspect that changes to the request or response characteristics are causing the increase. You need to identify whether the additional costs are driven by the model input size, the model output size, or expanded tool usage.Which observability capability should you use?
Answer(s): D
To identify whether model input, model output, or expanded tool usage is driving the cost spike, you should focus on token usage.Granular tracking of token usage breaks down the costs directly: Input vs. Output Tokens: Tracing token usage isolates whether the user prompt (input) has grown or if the model is generating longer, more expensive responses (output).Tool Calling: Expanded tool usage consumes extra tokens each time the agent schema processes an execution loop, which will clearly reflect as inflated token counts.
https://www.mindstudio.ai/blog/subtraction-principle-agent-harness-optimization
HOTSPOT (Drag and Drop is not supported)You have a Microsoft Foundry project that contains an agent.The agent uses tools to retrieve internal content and call external APIs. The agent is configured to let the model decide when to call the tools.You need to publish the agent for a compliance workflow. The solution must meet the following requirements:-Each workflow run must include a retrieval step before generating a response.-Tool calls must authenticate by using the published agent’s own identity.-Tool access must use an identity isolated from other project resources.-Tool access must use support audit tracing.What should you do? To answer, select the appropriate options in the answer area.Note: Each correct selection is worth one point.Hot Area:
Box 1: required Tool choiceTo comply with your strict workflow and compliance requirements, you set tool_choice parameter required.Setting tool_choice to required mathematically forces the model to invoke at least one tool (your internal retrieval mechanism) during the turn before it can finalize its answer.Incorrect [Not auto] Setting tool_choice to auto hands over absolute decision-making power to the underlying foundational model. If the model determines it already "knows" the answer from its training data, it can skip the retrieval step entirely.Box 2: Using a distinct agent identity bound to the client application Configure the tool to authenticate by Restriction: Use Published Agent's Own Identity Map the agent to a Microsoft Entra Agent Identity. This gives the published agent its own distinct service principal identity within Entra ID, ensuring it does not inherit or use user tokens or secondary administrative keys.Restriction: Identity Isolated from Project Resources Rather than sharing the default Microsoft Foundry project managed identity (project-mi), assign a unique User-Assigned Managed Identity exclusively to this agent instance. Limit its role-based access control (RBAC) scopes solely to the specific retrieval indexes and the target external API endpoints.
https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/tool-catalog
You have a Microsoft Foundry project named Project1 that contains the following:-An OpenAPI tool that calls an external API-A project connection named Connection1 that stores the API key of the external APIWhen an agent calls the OpenAPI tool, the API returns a 401 unauthorized error, and traces show that the API key header is NOT being sent.You need to ensure that the OpenAPI tool automatically includes the API key from Connection1 on all requests.What should you do?
You have a Microsoft Foundry project that contains a customer support agent. The agent calls an internal knowledge API tool before generating responses.Users report the following issues:-Some requests take more than 15 seconds to complete.-Some responses are incorrect, even when the knowledge API returns the expected data.You need to inspect individual agent runs to view the ordered sequence of large language model (LLM) calls, tool invocations, and timing information.Which observability capability should you use?
Here is LLM tracing (also known as trace view or distributed tracing for GenAI) needed.Tracks Execution Flow: It captures the exact ordered sequence of LLM calls and tool invocations.Pinpoints Latency: It provides timestamps and durations for every individual step to catch the 15-second bottlenecks.Inspects Inputs/Outputs: It lets you see the exact payload sent to and from the knowledge API and the final LLM prompt to find out why the agent hallucinated or ignored the data.
https://coralogix.com/ai-blog/advanced-techniques-for-monitoring-traces-in-ai-workflows/
Note: This section contains one or more sets of questions with the same scenario and problem. Each question presents a unique solution to the problem. You must determine whether the solution meets the stated goals. More than one solution in the set might solve the problem. It is also possible that none of the solutions in the set solve the problem.After you answer a question in this section, you will NOT be able to return. As a result, these questions do not appear on the Review Screen.You have a multimodal AI generative model that accepts image uploads and uses extracted image text to generate responses.You discover that users can upload unsafe images and embed hidden instructions into images to manipulate the model.You need to implement controls to mitigate the risk.Solution: You configure a prompt shield for user prompts.Does this meet the goal?
Note: This section contains one or more sets of questions with the same scenario and problem. Each question presents a unique solution to the problem. You must determine whether the solution meets the stated goals. More than one solution in the set might solve the problem. It is also possible that none of the solutions in the set solve the problem.After you answer a question in this section, you will NOT be able to return. As a result, these questions do not appear on the Review Screen.You have a multimodal AI generative model that accepts image uploads and uses extracted image text to generate responses.You discover that users can upload unsafe images and embed hidden instructions into images to manipulate the model.You need to implement controls to mitigate the risk.Solution: You configure image moderation to block unsafe content before processing the images.Does this meet the goal?
Share your comments for Microsoft AI-103 exam with other users:
Question 23:Question 23 describes a multimodal model where users can upload unsafe images that could contain hidden instructions. The goal is to implement controls to mitigate this risk. Key points to understand
Keeping this site free takes real effort. We constantly battle automated scraping and unauthorized content copying. A quick account helps us protect the community and keep the site free.
To continue studying for your AI-103, please sign in or create a free account.