AI-103 Practice Questions: Implement information extraction solutions Domain

Correct answer (A): The symptoms point to retrieval-pipeline problems, not model size. Irrelevant citations usually indicate issues such as poor chunking or missing metadata, and newly uploaded files not appearing strongly suggests ingestion or index freshness problems. The best first step is to review extraction quality, chunk boundaries, metadata, and whether the search index is current.

Why the other options are wrong:
- Option B: Content filters address safety concerns, not retrieval relevance, stale content, or missing citations.
- Option C: Fine-tuning is not the best first action when the likely root cause is in retrieval and indexing.
- Option D: A longer system prompt does not fix bad extraction, poor chunking, missing metadata, or stale index content.

Correct answer (C): The chat experience needs grounded answers with citations over extracted contract content. Indexing extracted text plus metadata in Azure AI Search enables retrieval of the right source passages, and hybrid or vector-assisted retrieval is usually better than keyword-only search for natural-language questions. Summaries and a larger model do not replace retrieval from source content when evidence and citations matter.

Why the other options are wrong:
- Option A: Nightly summaries can omit important evidence and weaken citation quality because the chat layer should retrieve from source content.
- Option B: Keyword-only search can miss semantically relevant passages when users ask natural-language questions.
- Option D: A larger chat model does not replace retrieval when users need grounded answers over a document corpus.

Correct answer (B): Managed identity with RBAC is the preferred Azure-native design because it avoids embedded secrets and supports least-privilege access. The app can be granted only the minimal permissions it needs on Azure Storage and Azure AI Search, which is especially important when extracted content contains sensitive employee data.

Why the other options are wrong:
- Option A: Environment variables still require secret management and are not preferred when managed identity is available.
- Option C: Encryption at rest protects stored secrets, but the design still depends on managing secrets instead of using keyless Azure-native authentication.
- Option D: Subscription-wide contributor access violates least-privilege guidance.

Correct answer (B): In a regulated workflow, uncertain extracted values should not drive automated decisions. The right design is to apply confidence thresholds and send low-confidence results to an exception or human-review path before making approval decisions. Logging, larger models, and content filters do not make uncertain extracted values trustworthy.

Why the other options are wrong:
- Option A: Logging helps with auditing, but it does not stop incorrect values from affecting regulated decisions.
- Option C: A larger model does not make uncertain extracted values reliable enough for regulated automation.
- Option D: Content filters are safety controls; they do not validate extraction accuracy for business fields.

Correct answer (B): Layout analysis is the right choice when the application must preserve structure such as reading order, paragraphs, headings, tables, and selection marks. The stem explicitly says there is no fixed schema yet, so structured field extraction is premature. OCR alone would flatten much of the structure, and retrieval is for question answering rather than preserving the document's original organization for review.

Why the other options are wrong:
- Option A: OCR can extract text, but it does not preserve the structural relationships the reviewers need.
- Option C: Structured field extraction is better when you already know the exact fields to return, which this scenario does not.
- Option D: Retrieval can help answer questions later, but it does not preserve tables, reading order, or selection marks.

Correct answer (B): A production extraction workflow should separate deterministic extraction from downstream conversational access. The strongest sequence is to ingest content, extract text, structure, and fields, validate low-confidence results, store normalized outputs, index approved searchable content, and only then enable grounded Q&A. Starting with summarization or raw chat skips validation and weakens both compliance and grounding.

Why the other options are wrong:
- Option A: Beginning with summarization reduces determinism and makes both field capture and grounded retrieval less reliable.
- Option C: Sending raw files directly to a chat model skips structured extraction and validation, which is risky in a regulated workflow.
- Option D: OCR alone does not provide normalized business fields, and content filters do not replace extraction validation.

Correct answer (C): When exact fields are required for a high-impact workflow, prompt-only extraction is weaker than schema-driven extraction and validation. This scenario also involves account changes, so low-confidence or risky outputs should be routed for human approval. Structured extraction plus validation provides better control than a single generative prompt or layout-only processing.

Why the other options are wrong:
- Option A: A stronger model may still produce plausible-looking values, but prompt-only extraction is not the best choice for exact, high-risk fields.
- Option B: A chat assistant can help answer questions, but it does not provide validated field extraction for account updates.
- Option D: Layout analysis preserves structure, but it does not by itself produce validated business fields for account changes.

Correct answer (C): A good extraction architecture separates stages such as ingestion, extraction, normalization, validation, storage or indexing, and downstream orchestration. That design improves control, auditability, and exception handling. Treating the entire workflow as one prompt call reduces governance, makes validation harder, and weakens traceability.

Why the other options are wrong:
- Option A: A better prompt may improve output quality somewhat, but it does not create the validation and audit controls the auditors require.
- Option B: Chunking may help with document size issues, but it still leaves the core design as prompt-centric and weak on control and provenance.
- Option D: Keyword search is useful for retrieval, not as a replacement for an extraction pipeline that feeds line-of-business systems.

Correct answer (B): Once content has been extracted and normalized, indexing the text plus useful metadata in Azure AI Search improves later retrieval for grounded question answering. This supports the assistant while preserving the structured extraction outputs. Search is a downstream retrieval layer; it complements extraction rather than replacing deterministic extraction needs.

Why the other options are wrong:
- Option A: Retrieval is valuable for Q&A, but it should not replace an existing extraction pipeline when structured outputs are still required.
- Option C: Embeddings alone do not preserve the detailed text and source metadata needed for clear grounding and review.
- Option D: Deleting detailed extraction output reduces traceability and can weaken retrieval quality.

Correct answer (B): Layout analysis is the best first step because the rules engine depends on document structure, not just plain text. It preserves tables, selection marks, and reading order, which are critical for deterministic downstream processing. OCR-only extraction can read text but will not reliably preserve those relationships. Prebuilt invoice extraction targets invoice fields, and summarization would throw away the structure the rules engine needs.

Why the other options are wrong:
- Option A: OCR-only processing can recover text, but it does not preserve tables, selection marks, or reading order well enough for structural rule processing.
- Option C: Prebuilt invoice models are intended for invoice-specific normalized fields, not general checklist structure.
- Option D: A free-text summary is not appropriate when the downstream system needs deterministic structural elements from the page.