Microsoft AI-103: Implement information extraction solutions Practice Questions

Hybrid search combines lexical/keyword matching (precise for exact terms like part numbers) with vector/semantic search (captures conceptual similarity like "failure" and "malfunction"), satisfying both requirements in a single query. Option A would miss the semantic/conceptual matches entirely. Option B risks missing exact-match precision that keyword search reliably provides for specific identifiers like part numbers. Option D describes a content-processing step, not a query-time search capability, and does not itself return ranked results.

Time-aligned transcription — indexing transcript text together with per-segment timestamps — enables both full-text search over spoken quotes and the ability to jump to the exact moment in the video where a match occurs. Option A loses the timestamp alignment needed to jump to the right moment, satisfying only the search half of the requirement. Option D provides essentially no searchable content related to what was actually said. Option C applies a visual capability that does not capture spoken quotes at all.

Semantic (vector) search matches based on conceptual meaning rather than requiring exact keyword overlap, which is exactly what's needed to surface relevant passages for a conceptual question phrased differently than the source text. Option D would likely return no useful results if the exact wording doesn't appear in the documents, failing the stated requirement. Option B is irrelevant — the query is already text, and speech-to-text does not affect search relevance. Option C is unrelated to the conceptual matching gap described.

Supporting both a scheduled batch pattern (for nightly PDF uploads) and event-driven/incremental updates (for near-real-time wiki edits) lets the index stay consistent with each source's natural update cadence, which is the correct design consideration. Option D would make wiki edits stale for up to a day, defeating the near-real-time requirement explicitly stated. Option B excludes an entire valid content source from the knowledge base. Option A is wasteful and unnecessary — a full rebuild on every small edit does not scale and is not required when incremental updates are viable.

A well-designed ingestion pipeline supports documents, images, audio and video, applying format-appropriate processing (OCR for scanned documents, transcription for audio, visual analysis for images) before indexing everything into a searchable store. Option A and D each wrongly exclude entire valid content categories that ingestion pipelines are designed to handle. Option C describes a manual process that defeats the purpose of an automated ingestion pipeline and does not scale to the described mixed-format corpus.