matsui528 / faiss_tips Goto Github PK

Firstly, thank you for your faiss tips and survey (A Survey of Product Quantization).
I have questions about HNSW + IVFPQ.

1. We use IFVPQ to encode vector (reduce memory) and HNSW as Coarse-quantization (reduce search time) ?
2. If we search on Coarse-quantization (HNSW), why we set index.nprobe (belong to IFVPQ) as parameter instead of index.hnsw.efSearch (belong to HNSW) ?
   Thank in advance.

Hi,

Could you please let me know if there is a way to update trained indexes with an incoming new data? I am particularly interested in deploying the LSH index.

In the section std::vector to/from np.array, index.xb is used but this attribute no longer exists.

A newer example using index.get_xb() works (facebookresearch/faiss#2266 (comment))

However I'm not clear why the 2nd parameter to faiss.rev_swig_ptr can be any number. It looks like gibberish values after n * d meaningful vector values are seen in the first n * d array positions.

I am struggling with how to assign parameters for a ivf index, specially how to set the nlist and nprobe ?

Also, what is the index_factory ?