Supported index algorithms
For dense vectors¶
These algorithms are based on Faiss library:
FaissIndex.L2— squared Euclidean (L2) distance.
FaissIndex.IP— this is typically used for maximum inner product search. This is not by itself cosine similarity, unless the vectors are normalized.
FaissIndex-es are constructed in
Flat mode, i.e. they implement exhaustive (precise) search.
To use Faiss implementation of HNSW index provide
For sparse vectors¶
These algorithms are derived from PySparNN library:
SparnnIndex.Cosine— Cosine distance (i.e.
1 - cosine_similarity).
SparnnIndex.JaccardBinary— Jaccard distance for binary vectors (i.e. vectors whose coordinates have the values 0 or 1).