Rust SDK#

USearch for Rust#

Detailed symbol list and documentation for USearch Rust SDK can be found on


cargo add usearch

This will add a USearch dependency to your Cargo.toml file.

usearch = "..."

By default, SimSIMD is used to provide dynamic dispatch for SIMD operations. You can, however, override that by specifying custom features in your Cargo.toml file. To disable all features, use the following configuration:

usearch = { version = "...", default-features = false }

To enable specific features, use the following configuration:

usearch = { version = "...", features = ["simsimd", "openmp", "fp16lib"] }

OpenMP (openmp) will use the OpenMP runtime for parallelism. It may not be available on all platforms, but on Linux it will lead to better performance and lower latency of small-batch operations on multi-core CPUs. The fp16lib flag will bring in the C-layer fp16 library to emulate half-precision floating point operations on older CPUs, where it may not be natively supported.


use usearch::{Index, IndexOptions, MetricKind, ScalarKind, new_index};

let options = IndexOptions {
    dimensions: 3, // necessary for most metric kinds
    metric: MetricKind::IP, // or MetricKind::L2sq, MetricKind::Cos ...
    quantization: ScalarKind::F16, // or ScalarKind::F32, ScalarKind::I8, ScalarKind::B1x8 ...
    connectivity: 0, // zero for auto
    expansion_add: 0, // zero for auto
    expansion_search: 0, // zero for auto

let index: Index = new_index(&options).unwrap();

assert!(index.capacity() >= 10);
assert!(index.connectivity() != 0);
assert_eq!(index.dimensions(), 3);
assert_eq!(index.size(), 0);

let first: [f32; 3] = [0.2, 0.1, 0.2];
let second: [f32; 3] = [0.2, 0.1, 0.2];

assert!(index.add(42, &first).is_ok());
assert!(index.add(43, &second).is_ok());
assert_eq!(index.size(), 2);

// Read back the tags
let results =, 10).unwrap();
assert_eq!(results.keys.len(), 2);


To save and load the index from disk, use the following methods:


Viewing the index does not load the data into memory, but allows you to inspect and traverse the index structure from external memory using memory-mapping. Similarly, serializing to/from in-memory buffers is supported. So you can memory-map the index file manually, and later call view_from_buffer or one of its siblings.

assert!(index.save_to_buffer(&mut serialization_buffer).is_ok());


USearch comes pre-packaged with SimSIMD, bringing over 100 SIMD-accelerated distance kernels for x86 and ARM architectures. That includes:

  • MetricKind::IP - Inner Product metric, defined as IP = 1 - sum(a[i] * b[i]).

  • MetricKind::L2sq - Squared Euclidean Distance metric, defined as L2 = sum((a[i] - b[i])^2).

  • MetricKind::Cos - Cosine Similarity metric, defined as Cos = 1 - sum(a[i] * b[i]) / (sqrt(sum(a[i]^2) * sqrt(sum(b[i]^2))).

  • MetricKind::Pearson - Pearson Correlation metric.

  • MetricKind::Haversine - Haversine (Great Circle) Distance metric.

  • MetricKind::Divergence - Jensen Shannon Divergence metric.

  • MetricKind::Hamming - Bit-level Hamming Distance metric, defined as the number of differing bits.

  • MetricKind::Tanimoto - Bit-level Tanimoto (Jaccard) metric, defined as the number of intersecting bits divided by the number of union bits.

  • MetricKind::Sorensen - Bit-level Sorensen metric.

User-Defined Metrics#

Custom metrics allow for the implementation of specific algorithms to measure the distance or similarity between vectors in the index. To use a custom metric with USearch, define a function that matches the expected signature and pass it to your index on creation, or later with change_metric. Let’s say you are implementing a weighted distance function to search through joint embeddings of images and textual descriptions of some products in a catalog, taking some UForm or CLIP-like models.

use simsimd::SpatialSimilarity;

let image_dimensions: usize = 768;
let text_dimensions: usize = 512;
let image_weights: f32 = 0.7;
let text_weights: f32 = 0.9;

let weighted_distance = Box::new(move |a: *const f32, b: *const f32| unsafe {
    let a_slice = std::slice::from_raw_parts(a, image_dimensions + text_dimensions);
    let b_slice = std::slice::from_raw_parts(b, image_dimensions + text_dimensions);

    let image_similarity = f32::cosine(a_slice[0..image_dimensions], b_slice[0..image_dimensions]);
    let text_similarity = f32::cosine(a_slice[image_dimensions..], b_slice[image_dimensions..]);
    let similarity = image_weights * image_similarity + text_weights * text_similarity / (image_weights + text_weights);

    1.0 - similarity

You can always revert back to one of the native metrics by calling:


Filtering with Predicates#

Sometimes you may want to cross-reference search-results against some external database or filter them based on some criteria. In most engines, you’d have to manually perform paging requests, successively filtering the results. In USearch you can simply pass a predicate function to the search method, which will be applied directly during graph traversal.

let is_odd = |key: Key| key % 2 == 1;
let query = vec![0.2, 0.1, 0.2, 0.1, 0.3];
let results = index.filtered_search(&query, 10, is_odd).unwrap();
    results.keys.iter().all(|&key| key % 2 == 1),
    "All keys must be odd"

Quantization and Custom Scalar Types#

USearch supports the Rust-native f32 and f64 scalar types, as well as the i8 for quantized 8-bit scalars. Goign beyond that, USearch supports non-native f16 and b1x8 for half-precision floating point and binary vectors, respectively.

Half Precision Floating Point#

Rust has no native support for half-precision floating-point numbers, but USearch provides a f16 type. It has no advanced functionality - it is a transparent wrapper around i16 and can be used with half or any other half-precision library. Assuming USearch uses the IEEE 754 no conversion is needed, you can unsafe-cast the outputs of other IEEE-compliant libraries to usearch::f16.

use usearch::f16 as USearchF16;
use half::f16 as HalfF16;

let vector_a: Vec<HalfF16> = ...
let vector_b: Vec<HalfF16> = ...

let buffer_a: &[USearchF16] = unsafe { std::slice::from_raw_parts(a_half.as_ptr() as *const SimF16, a_half.len()) };
let buffer_b: &[USearchF16] = unsafe { std::slice::from_raw_parts(b_half.as_ptr() as *const SimF16, b_half.len()) };

index.add(42, buffer_a);
index.add(43, buffer_b);

Binary Vectors#

USearch also implement binary distance functions and natively supports bit-vectors. If you initialize the index with quantization: ScalarKind::B1, you can add floating-point vectors and they will be quantized mapping positive values to 1 and negative and zero values to 0. Alternatively, you can use the b1x8 type to represent packed binary vectors directly.

let index = Index::new(&IndexOptions {
    dimensions: 8,
    metric: MetricKind::Hamming,
    quantization: ScalarKind::B1,

// Binary vectors represented as `b1x8` slices
let vector42: Vec<b1x8> = vec![b1x8(0b00001111)];
let vector43: Vec<b1x8> = vec![b1x8(0b11110000)];
let query: Vec<b1x8> = vec![b1x8(0b01111000)];

// Adding binary vectors to the index
index.add(42, &vector42).unwrap();
index.add(43, &vector43).unwrap();

let results =, 5).unwrap();

// Validate the search results based on Hamming distance
assert_eq!(results.keys.len(), 2);
assert_eq!(results.keys[0], 43);
assert_eq!(results.distances[0], 2.0); // 2 bits differ between query and vector43
assert_eq!(results.keys[1], 42);
assert_eq!(results.distances[1], 6.0); // 6 bits differ between query and vector42

Performance Tuning#

To optimize the performance of the index, you can adjust the expansion values used during index creation and search operations. Higher expansion values will lead to better search accuracy at the cost of slightly increased memory usage, but potentially much higher search times. Following methods are available to adjust the expansion values:

println!("Add expansion: {}", index.expansion_add());
println!("Search expansion: {}", index.expansion_search());

Optimizing hardware utilization, you may want to check the SIMD hardware acceleration capabilities of the index and memory consumption. The first will print the codename of the most advanced SIMD instruction set supported by the CPU and used by the index. The second will print the memory usage of the index in bytes.

println!("Hardware acceleration: {}", index.hardware_acceleration());
println!("Memory usage: {}", index.memory_usage());