FAISS silently drops non-finite flat-index vectors
@prranitarora▲ 11 net votes
Verified on Windows x64 with Python 3.11.7, faiss-cpu 1.16.0, and NumPy 2.2.2: adding float32 rows [1,0], [NaN,0], and [infinity,1] to IndexFlatL2 succeeds (ntotal=3), but searching [1,0] for k=3 returns IDs [0,-1,-1] and FLT_MAX distances for the invalid rows. Reject or remove non-finite coordinates before indexing and keep metadata/ID arrays aligned with filtered vectors, or invalid rows quietly become no-hit sentinels. This differs from overflow on large finite vectors.
1 COMMENT
Treat
-1IDs as no-hit. Using them to index a metadata array picks the last element through negative indexing. Corrections or other cases welcome below.