This package contains methods that can be used to replace big metric databases and its related metric function by a more light weight encoding and distance function. This package is designed to be used with SimilaritySearch.jl but can be used without it.

In particular, it contains the following methods/types:

• BinPerms: Binary encoding based on Brief permutations (with shift-based encoding)
• BinPermsDiffEnc: Binary encoding based on Brief permutations with differential encoding
• HyperplaneEncoding: Hyperplane-based binary encoding
• HighEntropyHyperplanes: Binary encoding based on high entropy hyperplanes
• NearestReference: String-based encoding using references
• Perms: Vector-based encoding of Permutations
• PCAProjection: Vector-based encoding just using PCA from MultivariateStats package.
• RandomProjection: Vector-based encoding using Gaussian random projections.

All methods use fit and predict methods with reasonable default parameters.

] add SurrogatedDistanceModels

using SimilaritySearch, SurrogatedDistanceModels

X = MatrixDatabase(rand(64, 10^5))
nbits = 256
refs = rand(X, 128)
B = fit(BinPerms, L2Distance(), refs, nbits)  # creates a BinPerms model mapping input vectors to `nbits`-bit vectors
binX = predict(B, X)  # projects the entire database `X` to the new Hamming space

## now using this to search
G = SearchGraph(; db=binX, dist=BinaryHammingSpace())
index!(G)

queries = MatrixDatabase(rand(64, 1000))
knns, dists = searchbatch(G, predict(B, queries), 10) # search queries for 10nn using the binary projection

TODO: Put citations for each method