Vector Databases Quiz

Answers and explanations appear as you take the quiz.

1. What should a vector database keep available alongside each indexed embedding?
   • Original content, a source identifier and relevant metadata
   • Only the numerical vector, discarding the source record
   • Only a similarity score computed at ingestion time
   • A model-generated answer with no link to the underlying item
2. An index uses embedding model A, and a new query service uses model B with the same vector length. What is the safe migration plan?
   • Mix both models' vectors because matching dimensions prove compatibility
   • Change only the distance threshold and leave stored vectors untouched
   • Build and evaluate a separately versioned model-B index before switching queries
   • Delete the original documents and retain the old embeddings
3. What tradeoff can approximate nearest-neighbor search make compared with exact search?
   • It proves each retrieved record is factually correct
   • It may gain speed or memory efficiency while losing some retrieval fidelity
   • It guarantees lower latency and identical recall on every collection
   • It removes the need to test relevance on representative queries
4. A search returns the smallest numerical vector distance in the collection. What does that distance establish?
   • The source document is verified as true
   • The user is authorized to read the document
   • The result is a calibrated probability of answer correctness
   • Only relative closeness under the chosen representation and distance function
5. A document has been removed from the source collection. What must the index workflow prevent?
   • Recomputing the document's metadata if it returns later
   • The old embedding continuing to make the removed document discoverable
   • New documents receiving different source identifiers
   • The query embedding being generated at request time
6. How should metadata filters be used when different users may access different documents?
   • Enforce each requesting user's permitted scope before showing results
   • Use the same broad collection for everyone, then hide the source link
   • Treat a high similarity score as proof of access rights
   • Apply filters only when the user asks for them explicitly
7. A new index makes queries faster but returns less useful passages. Which evaluation was missing?
   • Only a count of vectors stored per shard
   • Only a server latency benchmark on an empty collection
   • Only a check that both models emit numbers
   • Task relevance and retrieval fidelity on representative queries
8. What operational case should a vector-search design test before release?
   • Only a static index with no document updates
   • Only queries from an administrator account
   • Concurrent updates and fallback behavior when the index is unavailable
   • Only the model's embedding dimension printed in a log