Azure AI Apps and Agents Developer Associate · 33% of the exam

Implement generative AI and agentic solutions: free practice questions

5 sample questions from our 161-question bank for this domain — answers and explanations included. These are the same scenario-based style as the real Microsoft exam.

1. You need to ground an Azure AI Foundry agent with real-time enterprise data stored in Azure AI Search. The agent must dynamically retrieve relevant documents at query time rather than relying on pre-loaded file attachments. Which two actions should you take to achieve this? (Select TWO.)

  • A. Upload all Azure AI Search documents as file attachments and enable File Search on the agent
  • B. Create a custom function tool that accepts a search query and calls the Azure AI Search REST API, returning relevant results✓ Correct
  • C. Configure an Azure AI Search index as a connected data source and enable the Azure AI Search tool on the agent✓ Correct
  • D. Set the agent's system prompt to include the full contents of the Azure AI Search index at startup
  • E. Use Code Interpreter to directly query the Azure AI Search SDK within the sandboxed Python environment
Explanation

Options B and C are correct. Option B (custom function tool calling Azure AI Search) is a valid and flexible approach: the agent invokes the function, which queries the live index and returns fresh results at runtime—enabling true dynamic retrieval. Option C (configuring Azure AI Search as a connected data source with the native tool) is the first-class supported integration in Azure AI Foundry, providing direct, low-code grounding against a live index. Option A (uploading documents as file attachments) creates a static snapshot and does not reflect updates to the live index. Option D (embedding the full index in the system prompt) is impractical for large indexes due to context window limits and does not provide dynamic retrieval. Option E (Code Interpreter querying Azure AI Search SDK) is technically conceivable but the Code Interpreter sandbox has network restrictions and is not a supported or reliable pattern for enterprise data grounding.

2. A company is deploying a RAG solution using Azure AI Search. Their documents contain both structured metadata (product category, date, author) and unstructured body text. They want queries to leverage semantic understanding of the body text while also filtering results to a specific product category before ranking. Which combination of Azure AI Search capabilities achieves this?

  • A. Full-text search with OData filter expressions on the category field combined with the semantic ranker applied after filtering.✓ Correct
  • B. Vector-only search on the body text field, with post-query client-side filtering on the category field.
  • C. Full-text search only, using the category field as a scoring booster in a scoring profile.
  • D. Semantic ranker only, configured to apply the category filter as a semantic configuration field.
Explanation

Azure AI Search supports OData $filter expressions that are applied before ranking, efficiently narrowing the candidate set to a specific product category. The semantic ranker can then re-rank the filtered full-text or hybrid results using semantic understanding of the body text. This combination is the recommended pattern for mixed structured/unstructured queries. Vector-only search with client-side filtering is inefficient because all vectors must be retrieved and decoded before filtering, bypassing index-level optimizations. Using category as a scoring booster affects ranking scores but does not categorically exclude results outside the desired category. The semantic ranker does not accept OData filter expressions as part of its semantic configuration; filters are a separate query parameter applied at the engine level.

3. Your team is building an Azure AI Foundry agent that answers employee HR questions. The agent uses File Search to retrieve answers from uploaded HR policy PDFs. Users report that some answers contain outdated information from a previous policy version. The old policy files were deleted from the vector store, but the agent still returns stale answers. What is the most likely cause?

  • A. The agent's system prompt is cached and must be manually cleared after each policy update
  • B. The vector store associated with the File Search tool still contains embeddings from the deleted files✓ Correct
  • C. The Thread retains copies of file contents and must be deleted and recreated
  • D. The model's context window cached the old file contents in the current Run
Explanation

The vector store still containing embeddings from deleted files is the most likely cause. In Azure AI Foundry Agent Service, File Search uses a vector store that must have files explicitly removed and re-indexed; simply deleting a source file does not automatically remove its embeddings from the vector store. The system prompt is not cached in a way that would retain old document content; it contains instructions, not document text. Threads store message history but not full copies of indexed file contents for retrieval. The model's context window only spans a single Run and does not persist file contents across sessions.

4. A startup is building a real-time vision assistant that allows field technicians to photograph industrial equipment and ask natural-language questions about faults. They need a single model call that accepts both the image and a text question. Which Azure OpenAI model and API pattern should they use?

  • A. GPT-4o with the chat completions API, passing the image as a base64-encoded data URL or a publicly accessible HTTPS URL in the `image_url` content block of the user message✓ Correct
  • B. DALL-E 3 with the image generation API, including the technician's question as the prompt
  • C. `text-embedding-3-large` to embed the image, then perform a vector search query
  • D. GPT-4o-mini with a custom fine-tuned vision adapter trained on industrial fault images
Explanation

GPT-4o is a natively multimodal model that accepts image inputs via the chat completions API. Images are passed as `image_url` content blocks (base64 or HTTPS URL) alongside text in the user message, enabling a single API call for vision-language tasks. DALL-E 3 is an image generation model and cannot analyze or reason about input images. `text-embedding-3-large` produces text embeddings and does not accept image inputs for visual question answering. Fine-tuning GPT-4o-mini with a vision adapter is a complex, unverified approach that is not the standard pattern—GPT-4o vision capabilities are available out of the box without fine-tuning.

5. An Azure AI Foundry agent is configured with a Code Interpreter tool. During a Run, a user uploads a CSV file and asks the agent to clean the data and produce a new CSV. After the Run completes, the developer queries the Run steps and sees a `tool_call` step of type `code_interpreter` with status `completed`, but cannot find the output file. Where should the developer look to retrieve the generated output file?

  • A. In the Azure Blob Storage account linked to the Azure AI hub, under a container named after the Run ID
  • B. In the Run step's `tool_call` output, which contains a `file_id` referencing the file in the Files API; retrieve it using the Files endpoint✓ Correct
  • C. In the Thread's message list, where the agent automatically attaches generated files as message attachments with download URLs
  • D. In the agent's vector store, where Code Interpreter automatically indexes all generated files for future retrieval
Explanation

The Run step's tool_call output containing a file_id is correct. When Code Interpreter generates output files, the file references are embedded in the tool call output within the Run steps. The developer must parse the code_interpreter output for file_id values and then call the Files API endpoint to download the actual file content. Files are not automatically written to Azure Blob Storage linked to the hub; they are managed within the Agent Service's Files API. Generated files are not automatically appended as message attachments with download URLs — the developer must explicitly retrieve them from Run steps. Code Interpreter does not automatically index generated files into the vector store; vector stores are populated explicitly for File Search retrieval purposes.

156 more questions in this domain

Practice the full bank with instant grading, flashcards, and a timed mock exam.

Start practicing free