Retrieval-augmented generation has pushed hybrid search adoption hard this year. Enterprises building RAG pipelines increasingly pair vector search with full-text search and semantic ranking. They no longer treat either technique alone as good enough to ground large language model answers in real company data. That shift is exactly why AI-102 knowledge mining solutions skills matter right now. Specifically, the exam tests whether you can pick the right combination of indexing, enrichment, and retrieval for a given scenario.
This fourth set in our AI-102 series also lands right as an old API quietly runs out the clock. Azure AI Document Intelligence’s v2.0 API retires on August 31, 2026, which is barely two months from today. Therefore, question 32 below tests a migration scenario that’s now genuinely urgent, not just theoretical exam trivia.
This set covers the “Implement knowledge mining and document intelligence solutions” domain, worth 15-20% of the AI-102 exam. Specifically, you’ll work through indexers and skillsets, knowledge stores, and hybrid search with semantic ranking. In addition, you’ll cover custom document classification, custom extraction models, the Layout model, and Foundry IQ.
Each scenario mirrors the real exam’s format, too. A specific requirement rules out two or three otherwise-reasonable options, leaving one clearly correct answer. Read the scenario, choose your answer, then reveal the explanation to see exactly why the distractors fail. If you’re working through the series in order, start with AI-102 Questions 1-10 and Questions 21-30 first. This AI-102 knowledge mining solutions set assumes you’re already comfortable with the Foundry naming conventions from those posts.
One more thing worth flagging before question 31: Azure AI Search and Azure Document Intelligence solve genuinely different problems. This AI-102 knowledge mining solutions domain lumps them together, even so. Search finds and ranks existing content. Document Intelligence, in contrast, extracts structured fields from unstructured documents. As a result, several scenarios below hinge on picking the right service first, not the right feature within one service.
Question 31: Extracting Vendor Totals From Thousands of Invoices
Contoso’s finance team receives thousands of vendor invoices in dozens of different layouts every month. They want vendor name, invoice total, and line items extracted automatically, without training a custom model first.
A) Document Intelligence custom extraction model
B) Document Intelligence prebuilt invoice model
C) Azure AI Search skillset with OCR only
D) Azure AI Language key phrase extraction
👁 Reveal Answer
Correct Answer: B
Explanation: The prebuilt invoice model is designed exactly for this: vendor name, totals, and line items, with no training required. A custom extraction model would work eventually. However, it needs labeled training data first, so it’s more effort than this scenario calls for. Skillset OCR only pulls raw text, not structured invoice fields. Key phrase extraction identifies notable phrases in text, too, which isn’t the same as structured field extraction.
Question 32: Migrating Before an Imminent Retirement
Woodgrove Bank’s document pipeline still calls the Document Intelligence v2.0 API directly. A compliance review just flagged this dependency as a risk that needs a resolution date.
A) Leave it as is, since Document Intelligence versions don’t retire
B) Migrate to the current v4.0 (2024-11-30) API before August 31, 2026
C) Migrate to v3.0, since that’s the newest stable release
D) Wait until 2029 to migrate, since that’s the final retirement date
👁 Reveal Answer
Correct Answer: B
Explanation: Document Intelligence v2.0 retires on August 31, 2026. Therefore, migrating to the current v4.0 (2024-11-30) GA release now is the right call. Leaving it as is ignores a hard deadline that’s only about two months away at the time of this post. v3.0 isn’t the newest release, either. v4.0 is current, and v3.0 is itself already in its own retirement window, ending March 30, 2029. That 2029 date belongs to v3.0, not v2.0, so waiting that long would mean running an already-dead API for years.
Question 33: Building a Basic AI Enrichment Pipeline
Fabrikam has thousands of scanned PDFs sitting in Blob Storage. The team wants OCR, language detection, and key phrase extraction applied automatically. The output should end up searchable in an index afterward.
A) Azure AI Search indexer with an attached skillset
B) Document Intelligence custom classification model
C) Azure AI Language orchestration workflow
D) Azure Translator document translation
👁 Reveal Answer
Correct Answer: A
Explanation: An indexer paired with a skillset is exactly the AI enrichment pipeline pattern here. The indexer pulls documents from Blob Storage, and the skillset then runs OCR, language detection, and key phrase extraction. Afterward, the enriched output lands in a searchable index. Custom classification splits and identifies document types instead, which isn’t what’s being asked. Orchestration workflow routes conversational-language features, so it’s unrelated to document enrichment. Translator only translates text; it doesn’t build a search index at all.
Question 34: Saving Enriched Content for Data Science, Not Just Search
Northwind’s data science team wants the AI-enriched output of a skillset. Specifically, that means extracted entities and key phrases, saved as raw tables or files in storage. Critically, they need this separate from any search index.
A) Configure a knowledge store within the skillset
B) Enable semantic ranking on the search index
C) Attach a Shaper skill directly to the index, outside any skillset
D) Export the search index into Document Intelligence
👁 Reveal Answer
Correct Answer: A
Explanation: A knowledge store is secondary Azure Storage, tables or blobs. It holds a skillset’s enriched output for downstream data science or other non-search use. It’s defined inside the skillset itself, using projections. Semantic ranking reranks query results within a search index instead; it has nothing to do with exporting raw enriched content. A Shaper skill only exists inside a skillset, not attached directly to an index. Exporting a search index into Document Intelligence isn’t a real capability either way.
Question 35: Grounding a RAG Chatbot With Both Precision and Concepts
Contoso is building a RAG chatbot. It needs to match exact product codes and internal jargon precisely. At the same time, it must surface conceptually related content when there’s no exact keyword match at all.
A) Vector search only
B) Full-text keyword search only
C) Hybrid search combining vector and full-text queries with RRF
D) Semantic ranking applied to a vector-only index
👁 Reveal Answer
Correct Answer: C
Explanation: Hybrid search runs full-text and vector queries concurrently, then merges the results with Reciprocal Rank Fusion. That combination captures exact-match precision for jargon and product codes alongside vector search’s conceptual matching. Vector search alone can miss exact codes since it ranks by conceptual similarity, not exact terms. Full-text search alone misses conceptually related content when no keyword overlaps. Semantic ranker applies to textual content only. So, pairing it with a vector-only index skips the full-text side this scenario specifically needs.
Question 36: Improving Relevance of Hybrid Search Results
A hybrid search index already returns reasonable results through RRF. However, the team wants results reordered based on deeper language understanding of user intent, not merely the merged rank scores.
A) Add another vector field to the index
B) Enable the semantic ranker on the query
C) Increase the number of index replicas
D) Switch from HNSW to exhaustive KNN
👁 Reveal Answer
Correct Answer: B
Explanation: The semantic ranker reranks results using machine comprehension models that evaluate how well each result matches query intent. This layers right on top of RRF-merged hybrid results. Adding another vector field changes what’s searchable instead, not how results get reranked. Replicas affect capacity and query throughput, not relevance. Switching HNSW to exhaustive KNN changes the vector algorithm’s speed and exactness tradeoff; it doesn’t add language-based reranking either.
Question 37: Splitting and Classifying a Mixed Document Upload
A single uploaded PDF might contain a contract, an invoice, and a W-2 form. All three are merged into one file with no separators between them. The team wants each sub-document identified and split out automatically.
A) Document Intelligence Layout model
B) Document Intelligence custom classification model with document splitting
C) Azure AI Search skillset with key phrase extraction
D) Document Intelligence prebuilt invoice model only
👁 Reveal Answer
Correct Answer: B
Explanation: A custom classification model combines layout and language features. Specifically, it detects, splits, and identifies multiple document types bundled into one file. The Layout model extracts structure instead, meaning tables, styles, and paragraphs. However, it doesn’t classify or split distinct document types on its own. A skillset with key phrase extraction pulls terms from text instead, not document boundaries. The prebuilt invoice model only recognizes invoices, so it would miss the contract and W-2 content entirely.
Question 38: Extracting Fields From a Proprietary Internal Form
Litware has an internal form with a layout no prebuilt model supports. Unfortunately, the team only has a handful of sample scans available for training a model.
A) Document Intelligence prebuilt invoice model
B) Document Intelligence custom neural model trained on a single labeled sample
C) Azure AI Vision object detection
D) Azure AI Language custom text classification
👁 Reveal Answer
Correct Answer: B
Explanation: Custom neural models can train from as little as a single labeled sample, matching this low-data scenario. As a result, they adapt to a proprietary layout no prebuilt model covers. The prebuilt invoice model won’t recognize a proprietary form since it’s built for standard invoices specifically. Vision object detection identifies objects in photos, not text fields in scanned forms. Custom text classification sorts text into categories instead; it doesn’t extract positional, structured field values.
Question 39: Extracting Generic Structure From Any Document Layout
Adventure Works wants tables, selection marks, paragraph styles, and section-and-subsection hierarchy extracted from a wide variety of documents. Notably, they don’t want to define a fixed field schema first.
A) Document Intelligence Layout model
B) Document Intelligence prebuilt tax model
C) Azure AI Search knowledge store
D) Azure AI Language sentiment analysis
👁 Reveal Answer
Correct Answer: A
Explanation: The Layout model extracts text, selection marks, tables, styles, paragraphs, and hierarchical section-and-subsection structure generically, with no predefined schema required. The prebuilt tax model targets specific tax forms only, not general documents. A knowledge store just stores enriched output from a skillset; it doesn’t extract layout structure itself. Sentiment analysis scores emotional tone in text, which is unrelated to document structure entirely.
Question 40: Building One Reusable Knowledge Base Across Many Content Sources
Fabrikam has several Azure AI Search indexes spread across departments. The team wants one permission-aware knowledge layer that AI agents in the Microsoft Foundry portal can query. Critically, they don’t want to rebuild retrieval logic for each index separately.
A) Rebuild every index in Document Intelligence instead
B) Use Foundry IQ, the managed knowledge layer built on Azure AI Search
C) Enable a knowledge store on each individual skillset
D) Apply the semantic ranker to each index independently
👁 Reveal Answer
Correct Answer: B
Explanation: Foundry IQ is the managed knowledge layer that turns enterprise content into reusable, permission-aware knowledge bases for agents in the Microsoft Foundry portal. It’s built directly on top of Azure AI Search. Document Intelligence extracts document fields instead; it doesn’t serve as an agent-facing knowledge layer, so rebuilding there misses the point. A knowledge store per skillset only produces isolated enrichment output, not a unified layer. The semantic ranker improves relevance on individual indexes, too. However, it doesn’t unify multiple indexes into one reusable, permission-aware knowledge base.
Study Tips for AI-102 Knowledge Mining Solutions
A few habits will help with AI-102 knowledge mining solutions questions specifically. First, separate the two services cleanly in your head. Azure AI Search finds and ranks existing content, while Document Intelligence extracts structured fields from documents. Many scenarios hinge on this distinction before anything else.
Second, memorize the Document Intelligence retirement dates precisely. v2.0 retires August 31, 2026, v2.1 retires September 15, 2027, and v3.0 retires March 30, 2029. Mixing these up is an easy way to pick the wrong migration answer under time pressure.
Third, keep hybrid search and semantic ranker as two separate layers in your mind. Hybrid search merges vector and full-text results through RRF. In contrast, semantic ranker then reranks those merged results using language understanding, so it’s an addition, not a replacement.
Fourth, don’t confuse a knowledge store with Foundry IQ’s knowledge bases. A knowledge store holds one skillset’s enriched output for downstream analysis. In contrast, Foundry IQ unifies multiple search indexes into one permission-aware layer for agents.
Finally, read the Microsoft Learn overview of hybrid search in full. It also covers RRF and semantic ranking in more depth than any single exam question can.
Keep Practicing Your AI-102 Knowledge Mining Solutions Skills
That’s 40 questions down across four sets, spanning generative AI, computer vision, language processing, and now knowledge mining. Therefore, keep this AI-102 knowledge mining solutions momentum going. Together, these four domains cover the large majority of what the exam measures.
In addition, if you haven’t worked through them yet, start with AI-102 Questions 1-10 and Questions 21-30. These round out generative AI and language processing. Our AZ-305 mixed domain review is also worth a look if you’re studying multiple Microsoft certifications at once. Good luck, and see you in the next set.
