AI-103 Implement information extraction solutions: 169 practice questions

12 of the 169 Implement information extraction solutions questions in the Certsqill AI-103 bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for AI-103? Take the free 5-min readiness check →

1. Use integrated vectorization for indexing and queries: Which implementation best satisfies these requirements?

1. Apply semantic ranking to indexed text, then add query embeddings without rebuilding documents or checking vector dimensions.

   Document and query vectors require compatible dimensions and embedding space; changing representations requires appropriate reindexing or migration.
2. Use keyword retrieval with metadata filters, then apply semantic ranking to improve relevance for natural-language questions.

   Semantic ranking reranks text results but does not replace vector retrieval for semantic similarity across differently worded content.
3. Generate embeddings during ingestion and query execution, then use vector-only retrieval with document chunking and access metadata.

   Vector-only retrieval can support semantic similarity but does not provide keyword matching optimized for exact identifiers.
4. Use integrated vectorization for indexing and queries, then combine vector and keyword retrieval with metadata access filters. ✓

   This supports semantic similarity, exact identifier matching, and application access controls in one retrieval pipeline.

Integrated vectorization plus hybrid retrieval handles semantic and exact matching, while metadata filters enforce document access.

2. Populate the vector field by indexing compatible document: What should the developer do first?

1. Enable semantic ranking for the existing keyword results, then evaluate whether semantic questions retrieve additional documents.

   Semantic ranker can improve ordering or extract captions from an initial text or hybrid result set, but it does not generate embeddings or replace a missing vector field. This addresses reranking, not the missing vectorization prerequisite.
2. Add instructions telling the language model to infer semantic matches from retrieved keyword documents before generating grounded answers.

   Prompt instructions affect answer generation, not Azure AI Search candidate retrieval. They cannot populate vectors or make keyword results provide vector similarity, and retrieval should be diagnosed separately from generation.
3. Increase the number of retrieved keyword candidates, then rely on reciprocal rank fusion to create vector similarity signals.

   Hybrid fusion merges keyword and vector rankings; it does not create vector rankings when the vector side has no populated, comparable vectors. More keyword candidates cannot substitute for document and query embeddings.
4. Populate the vector field by indexing compatible document embeddings, then vectorize queries with the same embedding representation. ✓

   Hybrid retrieval requires comparable document and query vectors. The developer must generate compatible embeddings, populate the index, and ensure query vectorization uses the same model space and dimensions before evaluating vector matches.

The vector side of hybrid retrieval is unavailable because indexed documents lack compatible embeddings. Populate vectors and use compatible query vectorization first.

3. Combine keyword and vector retrieval: Which targeted correction best meets the requirements?

1. Replace vector retrieval with keyword retrieval, then apply semantic ranking to the filtered result set.

   Keyword retrieval can improve exact identifier matching, and semantic ranking can rerank candidates, but removing vectors risks reducing conceptual recall. It does not preserve the demonstrated semantic-retrieval strength.
2. Combine keyword and vector retrieval, merge their rankings, retain department filters, and evaluate candidate recall independently. ✓

   Hybrid retrieval combines keyword matching for exact identifiers with vector similarity for conceptual queries, then merges rankings. Metadata filters can enforce application access policy. Candidate recall should be evaluated separately because semantic reranking cannot recover excluded documents. Prerequisites include compatible vector dimensions and appropriate index configuration.
3. Shorten chunks, rebuild embeddings, retain vector-only retrieval, and rely on prompts to enforce department access.

   Shorter chunks can change retrieval granularity, but vector-only retrieval still lacks keyword matching for exact identifiers. Prompts are not an access-control mechanism; metadata filters must implement the application access policy.
4. Increase semantic-ranker priority, retain vector-only retrieval, and use captions to expose identifiers missed during retrieval.

   Semantic ranker reranks only the initial candidate set; it cannot recover identifiers whose documents vector retrieval excluded. Captions extract relevant source content but do not expand candidate retrieval.

Use hybrid retrieval with department filters, then evaluate candidate recall and semantic reranking separately.

4. Reindex with compatible embeddings: Which rollout step best satisfies these requirements?

1. Reindex with compatible embeddings, apply metadata access filters, and evaluate hybrid retrieval with identifier and semantic test queries. ✓

   This preserves vector compatibility, enforces access during retrieval, combines keyword matching with vector similarity, and makes quality measurable before release.
2. Keep the vectors, place access instructions in the prompt, and inspect generated answers for unauthorized document references.

   Prompt instructions do not implement application access policy. Retrieval eligibility must be restricted with metadata filters; answer inspection is not an adequate substitute.
3. Keep the vectors, enable semantic ranking, and compare generated captions with prototype answers before releasing the application.

   Semantic ranking reranks an initial candidate set; it does not repair missed exact identifiers in vector-only retrieval. Captions also do not replace retrieval evaluation.
4. Replace query embeddings with a newer model, retain the existing index, and measure semantic similarity before releasing the app.

   Query and document embeddings must use compatible model space and dimensions. Changing only query representations makes vectors incomparable, so documents must be reindexed or migrated.

Use compatible embeddings, metadata filtering, and hybrid retrieval, then evaluate exact-identifier and semantic queries separately.

5. Use one embedding model and vector dimension for indexing: What should the developer do first?

1. Use one embedding model and vector dimension for indexing and queries, then regenerate the indexed document vectors. ✓

   Document and query embeddings must be compatible in model space and dimensions. Reindexing creates document vectors matching the query representation.
2. Enable semantic ranking on the existing index, then rerank candidates without changing embeddings or rebuilding document vectors.

   Semantic ranker reranks an existing text or hybrid candidate set; it does not generate compatible embeddings or repair mismatched vectors.
3. Increase chunk overlap during indexing, then retain the current embedding representation and existing document vectors.

   Chunking affects retrieval granularity and context preservation, but changing overlap without regenerating compatible embeddings leaves the diagnosed mismatch unresolved.
4. Add keyword fields to the index, then enable hybrid retrieval while retaining the current document and query vectors.

   Hybrid retrieval can improve exact-term matching, but it does not make incompatible document and query vectors comparable.

Use compatible embedding representations for documents and queries, then reindex the documents.

6. Configure hybrid retrieval with compatible embeddings: Which targeted correction best addresses all three find

1. Run semantic ranking over current keyword candidates, then use captions to improve answers without changing candidate retrieval.

   Semantic ranker can improve ordering and captions, but it cannot recover semantically relevant documents absent from keyword candidates or enforce access filtering.
2. Re-embed documents with a new model, retain the old vector fields and query vectors, and increase semantic ranking.

   Changing embedding representations requires compatible query and document spaces and appropriate reindexing; retaining incompatible vectors does not fix retrieval.
3. Add a keyword-only query clause, preserve the vector fields, and improve identifier matching without addressing semantic recall or access.

   Keyword matching helps exact identifiers, but removing vector retrieval does not address semantic questions or enforce document access.
4. Configure hybrid retrieval with compatible embeddings, then apply per-user metadata filters for permitted documents. ✓

   Hybrid retrieval combines keyword and vector results, while metadata filters restrict candidates according to the application access policy.

Use hybrid keyword-and-vector retrieval with compatible embeddings, and apply metadata filters for each user's authorized documents.

7. Version the embedding model: What should the developer implement?

1. Replace hybrid retrieval with vector-only search, retain current vectors, and apply metadata filters before supplying results to generation.

   Metadata filtering is appropriate, but retaining vectors from a different representation risks incompatible comparisons, and vector-only retrieval weakens exact identifier matching.
2. Version the embedding model, reindex documents, use hybrid retrieval with access filters, then evaluate retrieval and reranking separately. ✓

   This aligns document and query vectors, preserves exact-term matching, enforces access policy during retrieval, and separates candidate recall from semantic reranking evaluation.
3. Reindex documents with the new embeddings, retain vector-only retrieval, and instruct generation to hide unauthorized retrieved content.

   Reindexing addresses representation compatibility, but vector-only retrieval can miss exact identifiers, while prompting cannot substitute for access-policy filtering.
4. Enable semantic ranker on existing keyword results, capture captions, and deploy without reindexing or measuring candidate recall.

   Semantic ranker reranks only the existing candidate set; it cannot recover paraphrase-relevant documents excluded by keyword retrieval. Captions do not replace retrieval evaluation.

Reindex compatible vectors, combine keyword and vector retrieval, filter by access metadata, and evaluate retrieval stages independently.

8. Reindex documents and regenerate query embeddings: What should the developer do to restore semantic retrieval

1. Use keyword search for queries, then retain the existing document vectors for the semantic portion of retrieval.

   Keyword search can preserve exact-identifier matching, but retaining incompatible vectors does not restore semantic retrieval. The vector representation must be rebuilt consistently.
2. Add department metadata filters, then rely on hybrid ranking to resolve the incompatible vector dimensions.

   Metadata filters restrict eligible documents for access policy; they do not change embedding dimensions or make incompatible vectors comparable.
3. Reindex documents and regenerate query embeddings with the same compatible model and vector dimensions. ✓

   Reindexing recreates document vectors in the replacement model’s space, while query vectorization uses the same model and dimensions. Hybrid retrieval can then combine semantic vectors with keyword matches.
4. Enable semantic ranking on the existing index, then use its captions to improve compatibility between vector representations.

   Semantic ranker reranks an existing text or hybrid candidate set and produces captions; it does not generate embeddings or repair incompatible vector dimensions.

Reindex with the replacement embedding model and use that same compatible model for query vectorization; retain hybrid keyword retrieval.

9. Reindex documents with the selected model: Which implementation change best addresses the findings?

1. Enable semantic ranking over the existing vector results, then retain the changed query embedding model for subsequent searches.

   Semantic ranker can rerank an initial candidate set, but it does not generate embeddings or make incompatible vectors comparable.
2. Increase chunk size during the next indexing run, while continuing to use different embedding models for documents and queries.

   Chunking changes retrieval granularity, but different embedding representations still prevent reliable vector comparison.
3. Reindex documents with the selected model, vectorize queries identically, and add hybrid keyword-vector retrieval. ✓

   This aligns document and query embeddings, then combines semantic similarity with keyword matching for exact identifiers.
4. Add a prompt instruction to recognize identifiers from passages returned through the existing vector-only retrieval process.

   Prompting cannot replace keyword matching or repair incompatible document and query vector spaces.

Reindex with one compatible embedding representation, vectorize queries the same way, and add hybrid retrieval for exact identifiers.

10. Version the index with compatible document and query: Which rollout step best satisfies these requirements?

1. Enable semantic ranking, use captions for answers, and rely on reranking to recover missed documents.

   Semantic ranker reranks an existing candidate set; it cannot recover documents excluded during initial retrieval. It also does not provide vector retrieval or enforce tenant access.
2. Rebuild with smaller chunks and vector-only retrieval while retaining access metadata fields.

   Compatible embeddings and chunking are valid considerations, but vector-only retrieval omits keyword matching for exact policy identifiers. Metadata fields alone do not apply access filters.
3. Version the index with compatible document and query vectorization, and combine keyword and vector retrieval with metadata access filters. ✓

   This implements hybrid retrieval: keyword matching supports exact identifiers, vectors support semantic similarity, and metadata filters enforce access policy. Versioning supports reproducible deployment.
4. Keep the current vectors, add keyword search, and ask the model to suppress unauthorized passages rather than filtering them during retrieval.

   Adding keyword search contributes to hybrid retrieval, but prompting the model to hide secrets is not an access-control mechanism. Metadata filters must restrict eligible documents before generation.

Use versioned hybrid retrieval with compatible vectorization and metadata filters.

11. Align the document and query embedding models: What should the developer do first to restore reliable retrieva

1. Enable semantic ranking and inspect captions to improve policy-question relevance.

   Semantic ranker reranks an existing candidate set and produces captions from source content. It does not generate embeddings or repair incompatible vector dimensions, so the underlying query failure remains.
2. Add keyword retrieval for identifiers and apply department filters to restrict document access.

   Keyword retrieval can help exact identifiers, and metadata filters can enforce document eligibility. However, this action does not repair the incompatible vector representations causing semantic-query failure.
3. Change chunking to preserve complete sections, then reindex without changing the embedding model.

   Chunking can balance context preservation and retrieval granularity, but reindexing with the unchanged incompatible representation does not align document and query vector spaces.
4. Align the document and query embedding models and dimensions, reindex the documents, and then evaluate hybrid retrieval. ✓

   Compatible model space and vector dimensions are required for meaningful comparison. Reindexing creates document vectors compatible with query vectors; hybrid retrieval can then combine semantic similarity with exact identifier matching.

Repair embedding compatibility and reindex before evaluating hybrid retrieval or semantic reranking.

12. Evaluate document freshness and answer support: What should the developer address next to reduce a remaining R

1. Evaluate document freshness and answer support, then require the app to state when evidence is insufficient. ✓

   Improved retrieval relevance does not establish that documents are current or that generated answers are supported. The application should inspect freshness and support, and acknowledge insufficient evidence instead of fabricating citations.
2. Increase chunk size across the index, then treat every retrieved passage as sufficient evidence for an answer.

   Chunk size is a retrieval trade-off, not proof that an answer is supported. Larger chunks may preserve context while reducing retrieval granularity, and retrieval still does not guarantee truth.
3. Change only query embeddings and compare them with the existing indexed vectors.

   Query and document embeddings require compatible model space and dimensions. Changing only query representations can make vectors incomparable; reindexing or migration is required when changing representations.
4. Replace hybrid retrieval with semantic ranking and use captions to supply missing policy details.

   Semantic ranker reranks an existing candidate set and captions extract source content; neither generates missing facts. Replacing hybrid retrieval can also weaken exact identifier matching.

Better retrieval does not guarantee current documents or supported answers; evaluate freshness and grounding, then handle insufficient evidence explicitly.

Other AI-103 domains

• Implement generative AI and agentic solutions — 429 questions →
• Plan and manage an Azure AI solution — 364 questions →
• Implement computer vision solutions — 169 questions →
• Implement text analysis solutions — 169 questions →
• All 1300 AI-103 questions →