About custom distance metric

[zineb198]

Hello,

Would it be possible to use the faiss package with user-defined metrics?

Thanks a lot for this package :D

[GioiaZheng]

FAISS does not support arbitrary user-defined (callback) metrics — search runs in C++/SIMD, so there's no hook for a Python distance function. The built-in metrics are METRIC_L2, METRIC_INNER_PRODUCT, and a set of extra ones: METRIC_L1, METRIC_Linf, METRIC_Lp (with metric_arg = p), METRIC_Canberra, METRIC_BrayCurtis, METRIC_JensenShannon, plus Jaccard/Hamming for binary indexes.

Practical workarounds:

• Reduce your metric to a built-in via preprocessing. Many "custom" metrics are L2/IP after a transform: cosine → L2-normalize the vectors and use METRIC_INNER_PRODUCT; Mahalanobis → apply the Cholesky/whitening transform (e.g. a LinearTransform) then METRIC_L2.
• Monotonic transforms preserve ranking. If your metric is a monotonic function of L2 or IP, the neighbor order is unchanged — search with the built-in and recompute exact scores only if you need the values.
• Coarse search + exact rerank. Use a built-in proxy metric to fetch top-k candidates, then rerank that small set with your exact custom metric in NumPy. This is the general escape hatch when the metric can't be transformed.