Implement generative AI and agentic solutions
Free Microsoft Azure AI Apps & Agents Developer (AI-103) practice — 6 questions on Implement generative AI and agentic solutions, with explanations. No sign-up. Full 12-question mixed test →
A healthcare company is building a multi-agent solution in Azure AI Foundry where a Triage Agent determines urgency, a Research Agent retrieves relevant medical literature, and a Response Agent formulates recommendations. The Research Agent occasionally returns irrelevant studies, causing the Response Agent to generate incorrect recommendations. Which architectural pattern BEST prevents the Response Agent from using low-quality retrieval results?
A Validator Agent acts as an orchestration safeguard between agents, explicitly evaluating intermediate outputs (retrieval results) before they propagate downstream. This is the recommended pattern for multi-agent workflows where one agent's output quality directly impacts another's performance. The validator uses a separate evaluation model to score relevance, filtering low-quality results before the Response Agent consumes them.
A financial services application uses a GPT-4o model deployed in Azure AI Foundry to generate investment summaries. The model occasionally fabricates company earnings data. The development team implements fabrication detection during evaluation and discovers the model hallucinates most frequently when summarizing companies with limited public data. Which combination of optimizations will MOST effectively reduce fabrications in production?
This combines three proven anti-hallucination strategies: (1) Phi-3 small models are specifically trained to be more factual and less prone to fabrication than larger models on knowledge-sparse queries, (2) RAG with SEC filings provides grounded, authoritative data that the model retrieves rather than invents, and (3) model reflection adds a self-verification step. For companies with limited public data, RAG ensures the model only summarizes what exists in verified sources.
A customer service agent built in Azure AI Foundry uses function-calling to check order status, process returns, and update account information. The agent frequently calls the update_account function even when customers only ask for information, causing unnecessary database writes and compliance audit flags. Token analytics show the agent is operating within budget. Which optimization will MOST effectively prevent inappropriate function calls while maintaining agent responsiveness?
This combines two critical safeguards for agent function-calling: (1) Explicit prompt guidance reduces model confusion about when to invoke functions—models often over-call functions without clear instructions, and (2) semiautonomous approval flows are the recommended pattern in Foundry for sensitive operations like database writes, giving users control before execution. This directly addresses the compliance concern while maintaining agent efficiency for read operations.
An enterprise RAG application in Azure AI Foundry connects to a Foundry project containing 50,000 internal policy documents indexed in Azure AI Search. Users report that queries about recent policy changes (last 30 days) often return outdated information from older versions of the same document. The search index was rebuilt 2 days ago. Latency breakdowns show retrieval takes 400ms and generation takes 1200ms. Which solution will MOST effectively ensure users receive current policy information?
AI Search boosting profiles with recency scoring is the native, performant solution for time-sensitive RAG scenarios. Exponential decay weighting gives progressively lower scores to older documents while still allowing them if they're highly relevant. Semantic ranking (L2 ranking) further improves relevance by understanding query intent. This is built into the search layer, requires no additional latency, and solves the problem at the retrieval source rather than post-processing.
A legal research agent built with Azure AI Foundry SDK must process complex queries requiring multi-step reasoning: (1) identify relevant statutes, (2) retrieve case law citing those statutes, (3) analyze precedent patterns, (4) formulate a legal argument. The current implementation uses a single GPT-4o call with chain-of-thought prompting, but responses lack depth in steps 3 and 4. Token analytics show average input tokens: 8000, output tokens: 2000, total: 10000 per query. Which workflow design will MOST improve reasoning quality?
Tool-augmented flows are the recommended pattern for multi-step reasoning where each step requires different data or processing. The agent can dynamically decide when to search statutes vs. retrieve cases vs. analyze patterns, adapting to query complexity. Conversation memory maintains context across tool calls without rigid sequential constraints. This gives the model agency to explore reasoning paths while keeping steps grounded in retrieved data, directly addressing the depth problem in steps 3-4 where analysis requires iterative exploration.
A manufacturing company's Azure AI Foundry agent monitors production metrics and automatically generates maintenance work orders when anomalies are detected. The agent has been in production for 3 weeks. Agent monitoring shows 127 work orders created, but only 43 resulted in actual maintenance (34% success rate). Operators report the agent creates orders for normal variations that don't require intervention. Tracing reveals the agent correctly identifies anomalies but over-estimates urgency. Which evaluation and optimization strategy will BEST improve the agent's precision?
Hybrid LLM plus rules-engine designs are the optimal pattern when some decisions have deterministic criteria and others need reasoning. Known failure patterns (e.g., temperature > X for Y minutes) should use rules for 100% precision and low latency. LLMs handle edge cases where context matters. This dramatically reduces false positives (the core problem) by removing LLM uncertainty from clear-cut cases, while keeping LLM flexibility for ambiguous scenarios. Operator feedback enables continuous evaluation of both components.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$109.99 $34.99 with code FREETEST33 — valid through September 2.
Get my $34.99 deal →