1Z0-184-25 Question 58
Single answerYou are designing a product recommendation engine on Oracle Cloud Infrastructure. Your application stores high-dimensional embeddings for product images and uses an IVF Vector Index to support approximate nearest neighbor (ANN) queries. As the dataset has grown, users report increased query latency, but you need to maintain strong recall (accuracy). Which approach is recommended to help reduce search time while preserving acceptable recall?
- A
Increase the number of centroid clusters in your IVF index so that each cluster contains fewer vectors, reducing the search space for each query.
- B
Disable partition filtering in your IVF index, forcing every cluster to be searched regardless of relevance.
- C
Store your index in Oracle Cloud Object Storage buckets instead of block storage to decrease retrieval time.
- D
Add more replicas of your index by copying it to multiple Availability Domains, ensuring each query is distributed.
Show answer and explanation
Correct answer: A
Explanation
When data volumes increase in IVF indexes, query speed can suffer if too many embeddings accumulate in each cluster. Adjusting the number of centroids (and tuning the parameters like the number of clusters searched) can strike a balance between recall and performance. For additional details on recommended practices and parameter tuning, consult Oracle Cloud Infrastructure documentation on vector index configurations.
- A. Correct.
Option 1: CORRECT. By increasing the number of centroid clusters in the IVF index, you narrow the candidate set assigned to each cluster. Since fewer vectors are examined during the search phase, query speed improves. You can still maintain good recall by selecting an optimal search parameter (e.g., searching a limited subset of clusters).
- B. Incorrect.
Option 2: INCORRECT. Disabling partition filtering (or cluster filtering) forces the system to check all clusters, which usually increases query latency, especially as the dataset grows. While this might theoretically improve recall, the performance trade-off is often unacceptable.
- C. Incorrect.
Option 3: INCORRECT. Simply storing your index in Object Storage rather than block storage generally increases latency because Object Storage is optimized for durability and cost rather than low-latency random I/O. It is not recommended for indexes requiring fast lookups.
- D. Incorrect.
Option 4: INCORRECT. While adding replicas can improve availability or failover, it does not necessarily solve the latency problem if each replica still has to process a large cluster size. Replication alone does not reduce the core challenge of large search spaces in each cluster.