Kuis Pencarian Multimodal

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. A text query retrieves an image. What makes that a multimodal search interaction?
   • Only images are indexed
   • The query must include a voice recording
   • Query and result use different forms of information
   • Every returned image is exact evidence
2. Two encoders output vectors with identical dimensions. What is still needed for meaningful comparison?
   • Only the same number of floating-point values
   • Training or alignment that makes the representations compatible
   • A shared file extension
   • A common UI color theme
3. Why combine keyword matching with embeddings for a product model number?
   • Keywords always capture visual shape better than images
   • Embeddings cannot rank any relevant results
   • Every model number has the same meaning
   • Exact identifiers can be lost by semantic similarity alone
4. A query asks for “a dog barking,” but retrieval returns a silent dog video. What failed?
   • The result matched visual subject but not required audio evidence
   • The dog was necessarily absent
   • The search system found no visual similarity
   • A text query cannot retrieve videos
5. Why might source metadata help a multimodal result?
   • An embedding always stores every legal right
   • Time range, origin and permissions support verification and use
   • Metadata proves every scene claim true
   • A thumbnail contains all video context
6. How should adding a second modality to ranking be evaluated?
   • Assume two modalities must beat one
   • Use only training similarity scores
   • On realistic relevance cases; more signals need not always improve results
   • Drop human relevance judgments
7. Which derived assets need consistent access controls?
   • Embeddings, thumbnails, transcripts and original files
   • Only the raw source file
   • Only the search box text
   • Only public result titles
8. An image result looks like the query item but is a different product. What should the team test?
   • Only average embedding length
   • Only the number of modalities in the database
   • Only whether all results have thumbnails
   • Fine distinctions among visually similar objects