1Z0-184-25 Question 18
Single answerYou are designing a product recommendation feature using embeddings stored in an Oracle Autonomous Database with vector columns. Your data science team has normalized all embeddings to unit length. However, they have been using Euclidean distance for similarity calculations and notice inconsistent recommendation results. Which distance function is most appropriate for this scenario to ensure consistent and semantically relevant search results?
- A
Cosine distance
- B
Euclidean distance
- C
Manhattan distance
- D
Jaccard distance
Show answer and explanation
Correct answer: A
Explanation
When embeddings are normalized to have unit length, cosine distance (or equivalently, maximizing cosine similarity) is often the preferred metric because it focuses on the angle between vectors rather than their magnitude. This approach is widely used in text embeddings, recommendation systems, and other AI-driven search scenarios. For further details, refer to Oracle Database vector column documentation, which provides guidelines on choosing distance metrics for vector searches.
- A. Correct.
Cosine distance (or cosine similarity) is typically used for vectors normalized to unit length. It effectively measures the angle between vectors and is a common metric for capturing semantic similarity in embedding-based scenarios.
- B. Incorrect.
Euclidean distance can be used for many vector-based searches, but it can produce misleading results when vectors are already normalized to unit magnitude, as their angular separation is more relevant than their absolute distance.
- C. Incorrect.
Manhattan distance measures the sum of absolute differences along each dimension and is rarely chosen for high-dimensional text or semantic embeddings, especially when dealing with unit-normalized vectors.
- D. Incorrect.
Jaccard distance is generally used for set-based comparisons, such as comparing sets of tokens, and is not best suited for continuous vector embeddings in semantic similarity searches.