1Z0-184-25 Question 55
Select 3Your company is developing a recommendation service on Oracle Cloud Infrastructure using an HNSW Vector Index for approximate nearest neighbor searches. The dataset is large, and new embeddings are added daily. You need to maintain both high recall and low query latency. Which three approaches align with recommended best practices for managing your HNSW Vector Index in this scenario?
- A
Periodically adjust the 'efSearch' parameter to optimize recall, with the understanding that higher values can increase query time.
- B
For significant daily additions, consider a full or partial rebuild of the index to maintain structure and performance, rather than relying solely on incremental updates.
- C
Use a direct 'AppendVectors' method that automatically re-optimizes the index structure with no need for periodic reindexing.
- D
Set the 'M' parameter to zero to reduce graph connections and shorten index creation time while maintaining high recall.
- E
Partition large datasets across multiple index shards or replicas for parallel queries and better scalability.
Show answer and explanation
Correct answers: A, B, E
Explanation
HNSW (Hierarchical Navigable Small World) indexing balances recall and performance through parameters like 'efConstruction', 'efSearch', and 'M'. In large-scale, frequently updated environments, it is common to periodically rebuild the index to maintain robust structure. Partitioning across multiple shards or replicas further improves query throughput. For in-depth guidance, refer to Oracle documentation on HNSW Index tuning and best practices for high-volume vector search use cases.
- A. Correct.
Correct. Increasing 'efSearch' heightens recall but can also raise query latency. Tuning it periodically ensures a suitable balance between performance and accuracy as your dataset grows.
- B. Correct.
Correct. If many new vectors are added at once, incremental updates alone might degrade index quality. Rebuilding the index for large batch additions is a standard practice to preserve accuracy and performance.
- C. Incorrect.
Incorrect. While incremental updates exist, there is no automatic 'AppendVectors' method that completely bypasses reindexing for large data additions. This distractor suggests an oversimplified solution that doesn't exist.
- D. Incorrect.
Incorrect. Setting 'M' to zero drastically impairs the connectivity in the graph, leading to poor recall. Proper tuning of 'M' is essential for effective indexing and query performance.
- E. Correct.
Correct. Sharding or replicating the index across multiple nodes can significantly improve query speed and scalability in large-scale vector search scenarios.