1Z0-184-25 Question 59
Single answerYour organization runs a recommendation system on Oracle Cloud Infrastructure that stores millions of product embeddings in a vector column within your Autonomous Database. You leverage an IVF (Inverted File) Vector Index to perform approximate nearest neighbor searches. Recently, you noticed that the recall rate of search results is dropping for queries on niche product embeddings. Which action should you take to enhance recall while maintaining acceptable performance with the IVF Vector Index?
- A
Decrease the number of cluster centroids (nlist) to reduce overall partitioning and set the number of probes (nprobe) to 1
- B
Increase the number of cluster centroids (nlist) and use a moderately higher number of probes (nprobe)
- C
Enable Hybrid Columnar Compression on the vector column to reduce the index size
- D
Create multiple parallel indexes on the vector column and combine their results at query time
Show answer and explanation
Correct answer: B
Explanation
To optimize IVF vector search, you need to balance centroid partitions (nlist) and the search probes (nprobe). Increasing nlist provides more refined clusters, improving the chance of hitting the correct cluster for niche embeddings. Meanwhile, adjusting nprobe (the number of clusters scanned during a query) helps maintain acceptable query times. Refer to Oracle� documentation on configuring vector indexes in Autonomous Database for best practices on tuning nlist and nprobe.
- A. Incorrect.
Option 1 is incorrect. Decreasing the number of centroids (nlist) while setting nprobe to 1 severely restricts the search space to only one cluster, leading to low recall for niche product embeddings. This helps performance but further reduces recall.
- B. Correct.
Option 2 is correct. When recall is too low, increasing the number of centroids (nlist) partitions the embedding space more finely, and using a moderate nprobe (like 10�20) searches across more clusters, improving recall without overly compromising performance.
- C. Incorrect.
Option 3 is incorrect. Hybrid Columnar Compression can reduce storage on certain data types but does not directly improve recall in vector searches. IVF performance tuning primarily revolves around the number of centroids and probes used.
- D. Incorrect.
Option 4 is incorrect. Creating multiple parallel vector indexes and merging their results adds unnecessary complexity, often without a tangible benefit in recall or performance. The recommended approach for tuning IVF indexes is adjusting the existing index parameters such as nlist and nprobe.