1Z0-184-25 Question 63
Single answerYou are building an IVF-based vector search index in Oracle Cloud Infrastructure for a dataset of more than two million product embeddings. Each vector has 768 dimensions, and your goal is to balance index creation time, memory usage, and search latency while achieving high recall for top-N results. Which approach is best suited for optimizing your IVF vector index in this scenario?
- A
Use a single centroid cluster (nlist = 1) to reduce indexing complexity and ensure the smallest memory footprint.
- B
Increase the number of centroid clusters (e.g., nlist = 2048 or higher) to maximize clustering granularity, ignoring the impact on indexing time and memory requirements.
- C
Employ an HNSW index exclusively, as IVF is not designed for large-scale vector searches.
- D
Use a moderately sized nlist (e.g., 256 or 512) and tune the search parameter nprobe to control the trade-off between latency and recall.
Show answer and explanation
Correct answer: D
Explanation
In Oracle Cloud Infrastructure, IVF (Inverted File) vector indexes are well-suited for handling large embedding datasets. By choosing a balanced number of clusters (nlist), you reduce index build time while still facilitating efficient approximate searches. The nprobe parameter can then be flexibly tuned to trade off between latency and recall. For more details on recommended configurations and best practices for vector indexes, consult Oracle� official documentation on OCI Vector Search.
- A. Incorrect.
Option 1 is incorrect because using only one centroid cluster (nlist = 1) eliminates the core benefit of IVF, which is to partition the data into multiple clusters to speed up approximate nearest neighbor search. With nlist = 1, you lose the ability to skip large portions of the dataset.
- B. Incorrect.
Option 2 is a common misconception. While increasing the number of centroid clusters (nlist) can improve recall by creating finer partitions, it also significantly increases index construction time, memory usage, and can lead to diminishing returns if set too high. It is not always the best choice in large-scale production scenarios.
- C. Incorrect.
Option 3 is incorrect because while HNSW is another ANN (Approximate Nearest Neighbor) method, it does not invalidate the usefulness of IVF for large datasets. IVF indexes can be highly effective when nlist and nprobe are tuned correctly, especially in scenarios that do not require the extreme recall and memory footprint of HNSW.
- D. Correct.
Option 4 is correct. Using a moderate number of clusters (nlist) and tuning nprobe allows you to strike a balance between index build time, memory consumption, and query latency. This approach leverages IVF� strengths for partitioning the dataset while providing flexibility to achieve high recall by adjusting the search parameters.