Task 1.1: Choose the appropriate Foundry services for generative AI and agents

• AA voice-based prompt agent connected through Voice Live

  Incorrect

  Voice-based prompt agents are for real-time spoken conversations. They are still declarative and don't run the team's code.

• BAn ephemeral agent that calls the Responses API from the team's own web server

  Incorrect

  Foundry wouldn't provide the endpoint, scaling or identity for code that runs on the team's own server.

• CA hosted agent that packages the existing code as a container or source bundle

  Correct

  Hosted agents run your own code and framework, such as LangGraph, as a container that Foundry hosts with a managed endpoint, scaling and a dedicated Entra identity.

• DA prompt agent defined by instructions, a model and a set of tools

  Incorrect

  A prompt agent is configuration only. It has no place for custom orchestration code or private Python libraries.

Pick the agent type by how much code you own. Instructions and tools only point to a prompt agent; existing framework code with custom dependencies points to a hosted agent, which Foundry runs for you.

• AAttach the web search tool to each agent and restrict it to the intranet domain

  Incorrect

  Web search grounds answers in public web results. It doesn't index private SharePoint or Blob content.

• BFine-tune the agents' model on the documents so that it can answer without retrieval

  Incorrect

  A fine-tuned model gives no citations and can't enforce who may see which document.

• CUpload the documents to a vector store for each agent and enable the file search tool

  Incorrect

  Copies of the documents in per-agent stores wouldn't carry SharePoint permissions, so access control would be lost.

• DCreate one Foundry IQ knowledge base over both sources and connect all three agents

  Correct

  A knowledge base can combine several knowledge sources, be shared by many agents, return answers with citations and enforce user permissions at query time.

Foundry IQ is the managed knowledge layer for this case: one knowledge base over several sources, reused by many agents, with agentic retrieval, citations and permission-aware results.

• ATwo deployments behind a classifier prompt that the app calls first to choose one

  Incorrect

  This works, but it is the routing code the team wants to avoid, and it adds a model call to every request.

• BModel router, set to a routing mode that favors cost

  Correct

  Model router is deployed like any model and chooses an underlying model per prompt. Balanced and Cost modes pick the most cost-effective model within a quality band.

• CA provisioned deployment of the current reasoning model

  Incorrect

  Provisioned throughput buys reserved capacity and steadier latency. It doesn't send simple prompts to a cheaper model.

• DA small language model for all prompts, with a larger maximum output length

  Incorrect

  Sending every prompt to a small model lowers cost but drops quality on the prompts that need reasoning; output length doesn't add reasoning.

Model router analyzes each prompt and routes it to a suitable model behind one deployment. Use Cost or Balanced mode to favor cheaper models, and compare it with your baseline in an evaluation before switching.

• AA large reasoning model for the intent labels, to get the best accuracy

  Incorrect

  For millions of short labels, a large model adds cost and latency that the task doesn't need.

• BA text embedding model for the photo questions

  Incorrect

  A text embedding model can't read the photo, and it doesn't produce an answer.

• CA multimodal model that accepts image input for the photo questions

  Correct

  Questions about what a photo shows need a model that reads images and text together.

• DAn image generation model for the photo questions

  Incorrect

  Image generation models create or edit pictures from prompts; they don't answer questions about an existing photo.

• EA small language model deployment for the intent labels

  Correct

  Labeling short messages is a narrow task. A small model handles it at lower cost and latency than a large model at this volume.

Match the model to the task: small models for narrow, high-volume text jobs; multimodal models when the input includes images. Image generation models create media rather than answer questions about it.

• AThe model's own training knowledge with a request to be current

  Incorrect

  Models have a knowledge cutoff, so they can't know this week's news.

• BThe web search tool, which grounds answers in current web results

  Correct

  Web grounding tools query the web for current information and cite their sources with URLs.

• CA file search vector store refreshed with news articles each month

  Incorrect

  Monthly uploads can't keep up with this week's events.

• DCode interpreter, asked to look up the news in Python code

  Incorrect

  Code interpreter isn't a search tool for current events.

For current public information, give the agent a web grounding tool; use file search and knowledge bases for your own content.

• AUpload the vendor's documentation to a vector store for file search

  Incorrect

  Documentation lets the agent read about the tools, not call them.

• BWrite a function tool that the app implements for each vendor action

  Incorrect

  That rebuilds by hand what the MCP server already exposes.

• CPaste sample tool requests into the agent's instructions

  Incorrect

  Instructions can describe actions but can't execute them.

• DAdd the MCP tool pointing at the vendor's remote MCP server endpoint

  Correct

  The MCP tool connects agents to remote MCP server endpoints and the tools they expose.

When a service already speaks MCP, connect it as an MCP tool instead of re-implementing its operations as custom functions.

• AThe model leaderboards and model comparison in the catalog

  Correct

  Leaderboards (preview) and side-by-side comparison show benchmark scores across quality, safety, cost and throughput, with trade-off charts.

• BA guardrail applied to each candidate model's deployment

  Incorrect

  Guardrails filter content at run time; they don't rank models.

• CThe AI gateway's token limits on each candidate model

  Incorrect

  Gateway policies govern traffic; they don't compare model quality.

• DDiagnostic logs from each model's deployment in production

  Incorrect

  Logs come after deployment and don't benchmark candidates.

Use benchmarks to shortlist, then evaluate the shortlisted models on your own data before deciding; public benchmarks don't replace tests on your task.

• AA scoring profile on the index that weights the main text field

  Incorrect

  Scoring profiles boost or suppress the ranking of documents already in an index; they don't load, chunk or embed any blobs.

• BSearch Explorer queries sent from the search service in the portal

  Incorrect

  Search Explorer runs queries against an index that already holds content; it doesn't build the ingestion pipeline.

• CA skillset with Text Split and embedding skills, written by hand

  Incorrect

  Those are the right skills, but writing the skillset, index and indexer definitions yourself is the JSON work the team wants to avoid.

• DThe Import data wizard in the Azure portal

  Correct

  The wizard chunks the content and calls the chosen embedding model to vectorize it, creating the data source, skillset, index and indexer.

For a quick start with integrated vectorization, use the portal wizard. The objects it creates have JSON definitions you can review and move into code or templates later, when the pipeline needs versioning.

• AThe azure-ai-inference package with a custom thread store

  Incorrect

  The azure-ai-inference package retired on August 26, 2026.

• BThe Assistants API, following the tutorial as written

  Incorrect

  The Assistants API sunset on August 26, 2026, so a new build shouldn't depend on it.

• CFoundry Agent Service, which is generally available

  Correct

  The Assistants API sunset on August 26, 2026; Microsoft points to the generally available Foundry Agent Service.

• DPrompt flow, with the agent built as a flow

  Incorrect

  Prompt flow is retiring and isn't recommended for new development.

Check lifecycle notices before building: the Assistants API and the azure-ai-inference package are retired, and prompt flow is retiring. New agents belong in Foundry Agent Service.