from scipy.sparse import random, vstack
from gensim.matutils import scipy2scipy_clipped
from sklearn.metrics.pairwise import cosine_similarity

#Some random sparse matrix
X = random(1000, 2000, density=.2, format="csc")

#Getting its similarity matrix
X_sim = cosine_similarity(X, dense_output=False)

#Splitting it to simulate chunking
X_sim_chunk1 = X[:500, :]
X_sim_chunk2 = X[500:, :]

#Assuring that in the first chunk no row is similar to the last item
X_sim_chunk1[:, -1] = 0

X_clipped1 = scipy2scipy_clipped(X_sim_chunk1, 100)
print(X_clipped1.shape) # (500, 1999)

X_clipped2 = scipy2scipy_clipped(X_sim_chunk2, 100)
print(X_clipped2.shape) # (500, 2000)

#While trying to recreate the matrix, this fails because of dimensions' inconsistency
vstack([X_clipped1, X_clipped2])
# ValueError: incompatible dimensions for axis 1