Faiss: How can we build index/search based on cosine similarity

[cherryPotter]

index = faiss.IndexFlatL2(d)
and
index.add(xb)
index = faiss.IndexIVFPQ(coarse_quantizer, d, nlist, m, faiss.METRIC_L2)

The above are all based on Euclid distance. How can I build index/search based on cosine similarity using faiss python package?

[bheinzerling]

index = faiss.IndexFlatIP(d)

IP stands for "inner product". If you have normalized vectors, the inner product becomes cosine similarity.

Here is an overview of the available indices:

https://github.com/facebookresearch/faiss/wiki/Faiss-indexes

[Berndinio]

This is searching for the cosine similarity! Not the cosine distance! Cosine distance is 1-.

[pranavnijampurkar33]

Just adding example if noob like me came here to find how to calculate the Cosine similarity from scratch

import faiss
dataSetI = [.1, .2, .3]
dataSetII = [.4, .5, .6]

dataSetII = [.1, .2, .3]

x = np.array([dataSetI]).astype(np.float32)
q = np.array([dataSetII]).astype(np.float32)
index = faiss.index_factory(3, "Flat", faiss.METRIC_INNER_PRODUCT)
index.ntotal
faiss.normalize_L2(x)
index.add(x)
faiss.normalize_L2(q)
distance, index = index.search(q, 5)
print('Distance by FAISS:{}'.format(distance))

To Tally the results check the cosine similarity of the following example

from scipy import spatial

result = 1 - spatial.distance.cosine(dataSetI, dataSetII)
print('Distance by FAISS:{}'.format(result))