IVF_FLAT is a coarse quantization index. When constructing the index, the vector dataset is clustered into nlist centroids based on similarity. During search, only the nprobe nearest clusters to the search vector need to be considered. The selection of the nprobe parameter involves a trade-off between recall rate and search efficiency. A larger nprobe leads to higher recall rates but lower efficiency.

Building parameters:

nlistNumber of cluster units[1,65536]
Table 9 Building Parameter of IVF_FLAT

Search parameters:

nprobeNumber of units to query[1, nlist]
Table 10 Search Parameter of IVF_FLAT