5 factors should be considered when selecting the Vector Search
Here are 5 factors that should be considered when selecting the Vector Search/Index algorithms.
For data sizes under 100K, a brute-force solution utilizing a FLAT index is sufficiently efficient.
Advanced algorithms may not offer significant speed improvements in such scenarios.
When an exact match is important then brute force is the right solution
Significant query latency reduction is achievable with a minor sacrifice in recall.
Some algorithms like HNSW are memory-hungry.
Scalar and Product Quantization significantly reduces storage consumption, at the expense of the Recall.
CPU vs GPU:
Usually moving from CPU to GPU provides a speed boost
Not all algorithms are optimized for GPU
Sometimes the building/indexing time is crucial.
IVF has a shorter indexing time compared to HNSW.