Anthropic’s Claude models just became generally available inside Microsoft Foundry on Azure. They run on Nvidia’s new GB300 Blackwell Ultra hardware. That launch is a good reminder of why AI-102 generative AI solutions skills matter right now. Enterprises are moving fast from AI experiments to production, and the exam tests exactly the design decisions those teams face every day.
This is the first set in our new AI-102 series, and every question below targets the same skill area. It focuses entirely on the “Implement generative AI solutions” domain, which is worth 15-20% of the exam. Specifically, you’ll work through RAG versus fine-tuning, hybrid search versus vector-only retrieval, content safety configuration, and multi-agent orchestration with Microsoft Foundry Agent Service. For example, several questions below mirror the exact tradeoffs teams are weighing as they move Claude, GPT, and other models into Microsoft Foundry deployments this year. Later posts in this AI-102 generative AI solutions series will branch into computer vision, language, and knowledge mining, but this first set stays focused on generative AI.
Each scenario mirrors the real exam’s format. A business constraint eliminates two or three plausible-sounding options, leaving one clearly correct choice. Read the scenario, pick your answer, then reveal the explanation to see why the distractors fail. As a result, you’ll build the same instinct the actual AI-102 exam rewards: matching a requirement to a service, not just recognizing a service’s name. Whether you’re new to Microsoft Foundry or migrating existing OpenAI workloads into it, these ten questions should sharpen your intuition before test day.
One more note before you dive in: Microsoft has been steadily renaming Azure AI services under the “Foundry” umbrella. Therefore, expect the real exam to use terms like “Microsoft Foundry Agent Service” and “Azure OpenAI in Foundry Models” instead of older names you might recall from study guides written a year or two ago. We’ve kept the terminology in this AI-102 generative AI solutions set current with what you’ll see on screen during the actual test.
Question 1: Default Content Filtering Behavior
Northwind Traders deployed a customer-support chatbot on Azure OpenAI in Foundry Models. The team used the out-of-the-box content filtering configuration. However, the compliance team now asks which severity level triggers blocking across the four harm categories: hate, violence, sexual, and self-harm. This applies to both prompts and completions.
A) Only content classified as High severity is blocked
B) Content classified at Medium or High severity is blocked; Low and Safe pass through
C) All four categories are blocked at any severity, including Low
D) Content is only annotated and never blocked unless a customer requests it
👁 Reveal Answer
Correct Answer: B
Explanation: The default configuration filters content detected at Medium or High severity. This applies to both prompts and completions, across all four harm categories. Low and Safe severity content passes through by default. Option A undersells the default policy, since Medium is also blocked. Option C describes the strictest configurable setting, not the default. Option D is wrong, too. Blocking happens automatically, not on request, though customers can still adjust the thresholds later.
Question 2: RAG vs. Fine-Tuning for a Policy Chatbot
Contoso’s legal department wants a chatbot that answers questions strictly from internal policy documents. Those documents change on a weekly basis. Answers must stay current, and they cannot hallucinate content that isn’t in the source material. However, the team has no budget for weekly retraining cycles.
A) Fine-tune GPT-4o on the policy documents every week
B) Implement a RAG pattern that grounds the model in the documents through Azure AI Search
C) Paste every policy document into the system message for each conversation
D) Enable groundedness detection instead of building a retrieval layer
👁 Reveal Answer
Correct Answer: B
Explanation: RAG grounds the model in an external, searchable index. Therefore, updating the documents updates the answers immediately, with no retraining required. Fine-tuning bakes knowledge into model weights instead. As a result, weekly retraining would be slow, costly, and still stale between cycles. Pasting entire documents into the system message also ignores token limits, and it breaks down fast as the document set grows. Groundedness detection only flags whether an existing response is supported by source material. It doesn’t provide retrieval on its own, so it can’t replace the RAG pipeline.
Question 3: Multi-Source Retrieval Without a Unified Index
Woodgrove Bank needs a knowledge assistant that searches SharePoint, Blob storage, and live web content together. As a result, the architecture team wants the LLM itself to break multi-part questions into targeted subqueries. It should also route each subquery to the right source, without the team building and maintaining one combined index.
A) Classic RAG using a single hybrid search index
B) Agentic retrieval with knowledge sources spanning multiple systems
C) Custom Question Answering with a static knowledge base
D) Semantic ranking applied to a single SharePoint indexer
👁 Reveal Answer
Correct Answer: B
Explanation: Agentic retrieval in Azure AI Search uses knowledge sources to unify multiple systems. In addition, it relies on LLM-assisted query planning to decompose complex questions into targeted subqueries automatically. Classic RAG with a single hybrid index requires you to build one unified index yourself. That is exactly what this scenario wants to avoid. Custom Question Answering fits static FAQ-style content, not multi-source live retrieval. Semantic ranking improves relevance within one index, but it doesn’t span multiple data sources or plan queries. In short, agentic retrieval is the only option built for this exact multi-source, LLM-planned pattern.
Question 4: Catalog Search That Needs Both Precision and Recall
Tailwind Traders’ product catalog serves shoppers who search in natural language, such as “waterproof jacket.” It also serves shoppers who search by exact SKU numbers, like SKU-4821-A. Vector search alone misses the exact SKU matches. Meanwhile, keyword search alone misses the synonym “waterproof” when a document says “water-resistant” instead.
A) Pure vector search using embeddings only
B) Pure keyword (BM25) search only
C) Hybrid search combining keyword and vector queries, merged with Reciprocal Rank Fusion
D) Semantic ranking applied without any vector fields
👁 Reveal Answer
Correct Answer: C
Explanation: Hybrid search runs keyword and vector queries in parallel. It then merges the ranked results with Reciprocal Rank Fusion. This captures both exact identifier matches and semantically similar phrasing. Pure vector search struggles with exact codes like SKU numbers, because embeddings don’t preserve them well. Pure keyword search, on the other hand, misses paraphrases such as “waterproof” versus “water-resistant.” Semantic ranking is a helpful second-stage reranker. However, it still needs an initial hybrid or vector query to rerank in the first place.
Question 5: Defending Against Indirect Prompt Injection
Litware’s support chatbot grounds its answers in customer-uploaded documents. Recently, attackers began embedding hidden instructions inside those files. Specifically, they hoped the model would follow the injected commands instead of answering the user’s actual question.
A) Raise the content filter severity threshold to Low for all categories
B) Enable Prompt Shields for indirect attacks, using document delimiters when constructing prompts
C) Turn off content filtering to reduce false positives
D) Fine-tune the base model on a handful of known attack examples
👁 Reveal Answer
Correct Answer: B
Explanation: Indirect attacks hide malicious instructions inside grounded data, such as an uploaded document, rather than in the user’s own prompt. Detecting them specifically requires document delimiters when embedding external content. That’s exactly what Prompt Shields for indirect attacks is built to catch. Raising the content filter severity targets harm categories like hate or violence instead, not prompt injection. Disabling filtering makes the system less safe, not more. Fine-tuning on a small attack sample won’t generalize well, either, since novel injection techniques keep changing.
Question 6: Real-Time Hallucination Detection in a Streaming Chat
Adventure Works already uses RAG for its streaming support chat. Now, the team wants real-time proof that each answer stays grounded in the retrieved documents. Specifically, they want to flag any response that drifts into unsupported claims.
A) Groundedness detection, available for streaming scenarios in select regions
B) The self-harm content filter category
C) Semantic ranking on the search index
D) Prompt flow evaluation run after deployment
👁 Reveal Answer
Correct Answer: A
Explanation: Groundedness detection flags whether an LLM’s response is actually supported by the source material provided to it. It is specifically designed for streaming scenarios, though it’s currently limited to a handful of supported regions. The self-harm filter, in contrast, addresses a completely unrelated harm category. Semantic ranking improves search relevance, not answer factuality. Prompt flow evaluation is valuable, too, but it’s a pre-production testing tool rather than a real-time, in-conversation check. Keep this distinction in mind: evaluation happens before launch, while groundedness detection runs live, during the conversation itself.
Question 7: Orchestrating Multiple Specialized Agents
Your team is building a workflow that needs several specialized agents collaborating autonomously on multi-step tasks. However, the team also wants minimal custom infrastructure code to maintain going forward.
A) Build a custom orchestration layer with Semantic Kernel from scratch
B) Use Microsoft Foundry Agent Service with Microsoft Agent Framework for multi-agent orchestration
C) Chain Azure Functions durable orchestrations with no agent framework
D) Rely on Custom Question Answering’s multi-turn conversation feature
👁 Reveal Answer
Correct Answer: B
Explanation: Microsoft Foundry Agent Service, combined with Microsoft Agent Framework, provides built-in support for complex workflows. This includes orchestration across multiple agents and users with autonomous capabilities. Building custom orchestration from scratch is possible, but it directly contradicts the “minimal custom infrastructure” requirement. Durable Functions can coordinate steps, too. However, they lack native agent-to-agent orchestration and reasoning. Custom Question Answering’s multi-turn feature handles conversational Q&A, not autonomous multi-agent collaboration.
Question 8: Preparing a Massive Document Set for Retrieval
Continuing this review, Fabrikam’s documentation team has amassed 10,000 pages of internal knowledge base articles. The chosen model accepts roughly 128,000 tokens. Therefore, sending every page with each request would waste tokens and degrade answer quality.
A) Request a custom model with an unlimited context window
B) Use Azure AI Search indexers and skillsets to chunk and vectorize the content for retrieval
C) Concatenate all 10,000 pages into a single oversized prompt
D) Search only by exact keyword match, skipping chunking entirely
👁 Reveal Answer
Correct Answer: B
Explanation: Indexers and skillsets automatically chunk large documents and generate embeddings. As a result, retrieval can pull only the most relevant passages instead of the entire corpus. An unlimited context window isn’t a real option. Even if it were, larger prompts still cost more and can dilute relevance. Concatenating everything into one prompt would exceed token limits immediately. Skipping chunking and relying on keyword search alone would also miss synonyms and paraphrased questions, which hurts recall.
Question 9: Comparing Prompt Flow Variants Before Launch
The architecture team at Contoso wants to compare three different prompt flow variants. Each variant uses a different model and prompt structure. The team needs a clear winner before choosing one for production.
A) Manually skim a handful of outputs after launch
B) Use Microsoft Foundry’s built-in evaluation tools to score each flow on relevance, groundedness, and related metrics
C) Deploy all three variants simultaneously and let customers decide
D) Use content filter logs as a proxy for output quality
👁 Reveal Answer
Correct Answer: B
Explanation: Microsoft Foundry includes evaluation capabilities built specifically for comparing models and flows. These tools score variants against metrics such as relevance and groundedness, before anything reaches production. Manually skimming outputs is subjective, and it doesn’t scale across variants. Deploying untested variants to customers risks a poor experience, too. That approach is not a substitute for structured evaluation. Content filter logs show safety events, not answer quality. Therefore, they can’t stand in for a real evaluation pass. Specifically, look for built-in metrics like coherence, fluency, and groundedness scores when you compare flows in Microsoft Foundry.
Question 10: Generating New Marketing Images from Text (AI-102 Generative AI Solutions)
Midway through this set, Woodgrove Bank’s marketing group needs brand-new promotional images. The team wants to generate them from text descriptions alone, with no existing photo to start from.
A) Azure Vision in Foundry Tools image analysis
B) The DALL-E model in Azure OpenAI
C) Azure AI Document Intelligence
D) Custom Vision object detection
👁 Reveal Answer
Correct Answer: B
Explanation: The DALL-E model in Azure OpenAI generates new images from text prompts. That’s exactly what a from-scratch marketing image request needs. Azure Vision in Foundry Tools analyzes existing images instead; it doesn’t create new ones. Document Intelligence extracts data from documents. Custom Vision classifies or detects objects in images you already have. Neither one generates original artwork, so both are ruled out here.
Study Tips for AI-102 Generative AI Solutions
A few habits make a real difference when you’re preparing for AI-102 generative AI solutions questions like these. First, map each scenario to a decision, not a feature list. The exam rarely asks what a service does. Instead, it asks which service wins under a specific constraint, such as cost, latency, or compliance.
Second, learn the current naming. Microsoft renamed several components to “Foundry” branding recently. So, know that Azure AI Search, Azure OpenAI in Foundry Models, and Microsoft Foundry Agent Service all now sit under the Microsoft Foundry umbrella. Older study guides may still reference “Azure AI Foundry” instead, so treat the two names as equivalent while you study.
Third, practice the RAG versus fine-tuning tradeoff until it becomes automatic. Use RAG when facts change often or must be traceable to a source. Consider fine-tuning only for stable style or format adjustments, and combine both approaches when a solution needs current facts plus a consistent voice.
Fourth, read the official documentation whenever you can’t fully picture a concept. For example, the Microsoft Learn overview of RAG in Azure AI Search is a strong reference for questions 2 through 4 above. Bookmark it, because RAG questions show up constantly across the generative AI domain.
Finally, time-box your review sessions. Twenty focused minutes on one domain beats two scattered hours across all four, especially once exam day gets close. Specifically, rotate between reading documentation and answering practice questions like these, since that mix tends to stick better than either one alone.
Keep Practicing Your AI-102 Generative AI Solutions Skills
Ten questions down, and plenty more ground to cover before exam day. Therefore, keep building on this AI-102 generative AI solutions foundation with the next set in the series. That next post moves into computer vision and natural language processing scenarios, which together make up nearly a third of the exam.
In addition, if you’re studying across multiple Microsoft certifications, our AZ-305 mixed domain review and AZ-305 cost optimization set are good companions. Our PL-200 renewal practice questions cover the Power Platform side of the Microsoft ecosystem, too. Come back tomorrow for the next set, and good luck with your studying.
