output
github_document

rsimsimd

{rsimsimd} is a light R wrapper around the rust crate simsimd. It allows efficient calculation of similarity metrics using SIMD operations

Installation

You can install the package like so:

# dev version
# install.packages("remotes")
# remotes::install_github("DavZim/rsimsimd")

# CRAN version
install.packages("rsimsimd")

Example

This is a basic example which shows you how to solve a common problem:

library(rsimsimd)

# a simple cosine similarity calculation
dist_cosine(c(1, 2, 3),
            c(4, 5, 6))
#> [1] 0.9746317

# more realistic embedding use case
# with 1536 (OpenAI) embedding dimensions
n_dimensions <- 1536
set.seed(123)
vec1 <- rnorm(n_dimensions)
vec2 <- rnorm(n_dimensions)

dist_cosine(vec1, vec2)
#> [1] -0.01064177

# if you have a list (or a vector) of embeddings (eg a database)
# and you want to compare a list of vectors, you can achieve this like so

# simulate a DB of 1000 embedding vectors
db <- lapply(seq(1000), function(x) rnorm(n_dimensions))

# simulate a lookup of 3 embedding vectors
lookup <- lapply(seq(3), function(x) rnorm(n_dimensions))

res <- dist_cosine(lookup, db)
# one row for each lookup, one column for each DB entry
dim(res)
#> [1]    3 1000

Functions

• [dist_cosine()] Cosine Similarity Matrix [ ] dist_dot [ ] dist_sqeuclidean [ ] div_jensenshannon [ ] div_kullbackleibler

• [get_capabilities()] reports current hardware capabilities

Benchmark

# alternative implementation of cosine, taken from lsa
# see also https://github.com/cran/lsa/blob/master/R/cosine.R
lsa_cosine <- function(x, y) {
  as.vector(crossprod(x, y) / sqrt(crossprod(x) * crossprod(y)))
}

set.seed(123)
n_dimensions <- 1536
v1 <- rnorm(N)
v2 <- rnorm(N)

bench::mark(
  lsa_cosine(v1, v2),
  dist_cosine(v1, v2)
)
#> # A tibble: 2 × 6
#>   expression               min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>          <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 lsa_cosine(v1, v2)    6.79µs   7.24µs   118470.        0B        0
#> 2 dist_cosine(v1, v2)   6.72µs    8.5µs    79961.        0B        0

# compare 1 embedding to 1'000 embeddings
ll_1k <- lapply(seq(1000), function(i) rnorm(n_dimensions))

bench::mark(
  sapply(ll_1k, function(ll) lsa_cosine(v1, ll)),
  dist_cosine(v1, ll_1k),
  check = FALSE # rounding errors
)
#> # A tibble: 2 × 6
#>   expression                             min median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                          <bch:> <bch:>     <dbl> <bch:byt>    <dbl>
#> 1 sapply(ll_1k, function(ll) lsa_cos… 7.94ms 8.52ms      112.    31.7KB     2.07
#> 2 dist_cosine(v1, ll_1k)              1.25ms 1.78ms      521.    35.6KB     0

# compare 1 embedding to 100'000 embeddings
ll_100k <- lapply(seq(100000), function(i) rnorm(n_dimensions))

bench::mark(
  sapply(ll_100k, function(ll) lsa_cosine(v1, ll)),
  dist_cosine(v1, ll_100k),
  check = FALSE # rounding errors
)
#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 2 × 6
#>   expression                           min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                      <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 sapply(ll_100k, function(ll) l…    1.07s    1.07s     0.932    3.29MB    0.932
#> 2 dist_cosine(v1, ll_100k)        170.37ms 178.28ms     5.67     5.96MB    0

# 1k x 1k comparisons => 1mln comparisons
bench::mark(
  dist_cosine(ll_1k)
)
#> # A tibble: 1 × 6
#>   expression              min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>         <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 dist_cosine(ll_1k)    299ms    307ms      3.26    7.66MB        0