uint40_t.maratyszcza/fp16.ashvardanian/simsimd. pip install usearch.npm install usearch.cargo add usearch.cloud.unum:usearch on GitHub.reserve.Most vector-search packages focus on just 2 metrics - "Inner Product distance" and "Euclidean distance". The lack of dedicated "Cosine distance" can be justified with the simplicity of normalizing such vectors on the fly. But that hardly exhausts the list of possible metrics.
Older approaches indexing high-dimensional spaces, like KD-Trees and Locality Sensitive Hashing are hardly extendible to vectors/objects of variable length. Modern NSW-like approaches, however, only require two objects to be comparable. This means, the index can be extended to support fuzzy search with Levenstein distance over strings. USearch implements HNSW, and provides some uncommon-to-vector-search "similarity measures" and "distances" out of the box: Haversine, Jaccard, Hamming, etc.
Most AI, HPC, or Big Data packages use some form of a thread pool.
Instead of spawning additional threads within USearch, we focus on thread-safety of the add function.
```cpp
for (std::size_t i = 0; i < n; ++i)
native.add(label, span_t{vector, dims}, omp_get_thread_num());
```
During initialization we allocate enough temporary memory for all the cores on the machine. On call, the user can simply supply the identifier of the current thread, making this library easy to integrate with OpenMP and similar tools.
Constructing a vector index is a memory-intensive task. Depending on your configuration parameters, part of the memory will be allocated towards copies of vectors, and another part to the Small World Graph, common to NSW-like algorithms. Choosing the right data-structures and numeric types can have profound implications on both.
Internally, some point "label" is mapped into an integer identifier for every stored vector.
Most implementations hard-code it to uint32_t, as people rarely store more than 4 Billion vectors.
But when they do, software becomes unusable.
Others hard-code it to uint64_t, which hardly space-efficient.
We added a tiny class - uint40_t, that should be enough for collections up to 1 Trillion vectors.
To construct a large index one may use a beefy RAM-optimized machine, like the AWS u-24tb1.metal instances with 24 TB of RAM.
Those are, however, pricy, and would cost you over $200/hour.
| | To Build | To Serve | | :------- | :-----------: | :------------: | | Instance | u-24tb1.metal | is4gen.8xlarge | | Price | ~ $200/h | ~$4.5/h | | RAM | 24 TB | 192 GB | | SSD | EBS | 30 TB |
You wouldn't want to have them running all the time. So, if you are working on large datasets, but don't need the in-RAM throughput, you can simply view an existing dataset from disk, without ever loading it fully into memory.
Construct on large machines, deploy on the cheap ones.
Most modern CPU have at least partial support for half-precision f16_t arithmetic.
USearch supports automatic down-casting and up-casting between f32_t, f16_t, f64_t, and i8q100_t representations.
Making vectors smaller will help pack more of them in-memory, while also increasing performance on CPUs implementing native support for target type.
Below are the performance numbers for a benchmark running on the 64 cores of AWS c7g.metal "Graviton 3"-based instances.
We fix the default configuration in the top line and show the affects of various parameters by changing one parameter at a time.
| Vectors | Connectivity | EF @ A | EF @ S | Add, QPS | Search, QPS | Recall @1 |
| :--------: | :----------: | :----: | :----: | :----------: | :-------------: | ------------: |
| f32 x256 | 16 | 128 | 64 | 75'640 | 131'654 | 99.3% |
| | | | | | | |
| f32 x256 | 12 | 128 | 64 | 81'747 | 149'728 | 99.0% |
| f32 x256 | 32 | 128 | 64 | 64'368 | 104'050 | 99.4% |
| | | | | | | |
| f32 x256 | 16 | 64 | 32 | 128'644 | 228'422 | 97.2% |
| f32 x256 | 16 | 256 | 128 | 39'981 | 69'065 | 99.2% |
| | | | | | | |
| f16 x256 | 16 | 64 | 32 | 128'644 | 228'422 | 97.2% |
| f32 x256 | 16 | 256 | 128 | 39'981 | 69'065 | 99.2% |
The main columns are:
To read more and reproduce, jump to benchmarking section.
There are two usage patters:
usearch/usearch.hpp, only available in C++.To use in a C++ project simply copy the include/usearch/usearch.hpp header into your project.
Alternatively fetch it with CMake:
cmake
FetchContent_Declare(usearch GIT_REPOSITORY https://github.com/unum-cloud/usearch.git)
FetchContent_MakeAvailable(usearch)
The simple usage example would require including the unum::usearch namespace and choosing the right "distance" function.
That can be one of the following templates:
cos_gt<float> for "Cosine" or "Angular" distance.ip_gt<float> for "Inner Product" or "Dot Product" distance.l2_squared_gt<float> for the squared "L2" or "Euclidean" distance.jaccard_gt<int> for "Jaccard" distance between two ordered sets of unique elements.bit_hamming_gt<uint> for "Hamming" distance, as the number of shared bits in hashes.pearson_correlation_gt<float> for "Pearson" correlation between probability distributions.haversine_gt<float> for "Haversine" or "Great Circle" distance between coordinates.That list is easily extendible, and can include similarity measures for vectors that have a different number of elements/dimensions. The minimal example would be.
```c++ using namespace unum::usearch;
index_gt
index.reserve(10); index.add(/* label: / 42, / vector: / {&vec, 3}); index.search( / query: / {&vec, 3}, / top / 5 / results /, / with callback: */ { });
index.save("index.usearch"); // Serializing to disk index.load("index.usearch"); // Reconstructing from disk index.view("index.usearch"); // Memory-mapping from disk ```
The add is thread-safe for concurrent index construction.
For advanced users, more compile-time abstraction are available.
cpp
template <typename metric_at = ip_gt<float>, //
typename label_at = std::size_t, // `uint32_t`, `uuid_t`...
typename id_at = std::uint32_t, // `uint40_t`, `uint64_t`...
typename scalar_at = float, // `double`, `half`, `char`...
typename allocator_at = std::allocator<char>> //
class index_gt;
One may also define a custom metric, such as Damerau–Levenshtein distance, to compute the similarity between variable length strings. The only constraint is the function signature:
cpp
struct custom_metric_t {
T operator()(T const* a, T const* b, std::size_t a_length, std::size_t b_length) const;
};
Python bindings are implemented with pybind/pybind11.
Assuming the presence of Global Interpreter Lock in Python, on large insertions we spawn threads in the C++ layer.
```python $ pip install usearch
import numpy as np import usearch
index = usearch.Index( ndim=256, # Define the number of dimensions in input vectors metric='cos', # Choose the "metric" or "distance", default = 'ip', optional dtype='f32', # Quantize to 'f16' or 'i8q100' if needed, default = 'f32', optional connectivity=16, # How frequent should the connections in the graph be, optional expansionadd=128, # Control the recall of indexing, optional expansionsearch=64, # Control the quality of search, optional )
vector = np.random.uniform(0, 0.3, (index.ndim)).astype(np.float32) index.add(42, vector) matches, distances, count = index.search(vector, 10) ```
Same can be done with batches, rather than single entries. Under the hood, worker threads will be spawned, to parallelize the procedure.
```py n = 100 labels = np.array(range(n), dtype=np.longlong) vectors = np.random.uniform(0, 0.3, (n, index.ndim)).astype(np.float32)
index.add(labels, vectors, copy=True) matches, distances, counts = index.search(vectors, 10) ```
```js // npm install usearch
var index = new usearch.Index({ metric: 'cos', connectivity: 16, dimensions: 2 }) assert.equal(index.connectivity(), 16) assert.equal(index.dimensions(), 2) assert.equal(index.size(), 0)
index.add(15, new Float32Array([10, 20])) assert.equal(index.size(), 2)
var results = index.search(new Float32Array([13, 14]), 2) assert.deepEqual(results.labels, new Uint32Array([15, 16])) assert.deepEqual(results.distances, new Float32Array([45, 130])) ```
Being a systems-programming language, Rust has better control over memory management and concurrency, but lacks function overloading.
Aside from the add and search, it also provides add_in_thread and search_in_thread which let users identify the calling thread to use underlying temporary memory more efficiently.
```rust // cargo add usearch
let quant: &str = "f16"; let index = new_ip(5, &quant, 0, 0, 0).unwrap();
assert!(index.reserve(10).isok()); assert!(index.capacity() >= 10); assert!(index.connectivity() != 0); asserteq!(index.dimensions(), 5); assert_eq!(index.size(), 0);
let first: [f32; 5] = [0.2, 0.1, 0.2, 0.1, 0.3]; let second: [f32; 5] = [0.2, 0.1, 0.2, 0.1, 0.3];
assert!(index.add(42, &first).isok()); assert!(index.add(43, &second).isok()); assert_eq!(index.size(), 2);
// Read back the tags let results = index.search(&first, 10).unwrap(); assert_eq!(results.count, 2);
// Validate serialization assert!(index.save("index.usearch").isok()); assert!(index.load("index.usearch").isok()); assert!(index.view("index.usearch").is_ok());
// There are more "metrics" available assert!(newl2(5, &quant, 0, 0, 0).isok()); assert!(newcos(5, &quant, 0, 0, 0).isok()); assert!(newhaversine(&quant, 0, 0, 0).isok()); ```
java
Index index = new Index.Config().metric("cos").dimensions(2).build();
float vec[] = {10, 20};
index.add(42, vec);
int[] labels = index.search(vec, 5);
```swift let index = Index.l2(dimensions: 4, connectivity: 8) let vectorA: [Float32] = [0.3, 0.5, 1.2, 1.4] let vectorB: [Float32] = [0.4, 0.2, 1.2, 1.1] index.add(label: 42, vector: vectorA[...]) index.add(label: 43, vector: vectorB[...])
let results = index.search(vector: vectorA[...], count: 10) assert(results.0[0] == 42) ```
All major HNSW implementation share an identical list of hyper-parameters:
M),efConstruction),ef).The default values vary drastically.
| Library | Connectivity | EF @ A | EF @ S |
| :-------: | :----------: | :----: | :----: |
| hnswlib | 16 | 200 | 10 |
| FAISS | 32 | 40 | 16 |
| USearch | 16 | 128 | 64 |
To achieve best results, please compile locally and check out various configuration options.
```sh cmake -B ./buildrelease \ -DCMAKEBUILDTYPE=Release \ -DUSEARCHUSEOPENMP=1 \ -DUSEARCHUSEJEMALLOC=1 && \ make -C ./buildrelease -j
./build_release/bench --help ```
Which would print the following instructions:
```txt
SYNOPSIS
./build_release/bench [--vectors
OPTIONS
--vectors
--queries <path>
.fbin file path to query the index
--neighbors <path>
.ibin file path with ground truth
-b, --big Will switch to uint40_t for neighbors lists with over 4B entries
-j, --threads <integer>
Uses all available cores by default
-c, --connectivity <integer>
Index granularity
--expansion-add <integer>
Affects indexing depth
--expansion-search <integer>
Affects search depth
--native Use raw templates instead of type-punned classes
--f16quant Enable `f16_t` quantization
--i8quant Enable `int8_t` quantization
--ip Choose Inner Product metric
--l2 Choose L2 Euclidean metric
--cos Choose Angular metric
--haversine Choose Haversine metric
-h, --help Print this help information on this tool and exit
```
AI has a growing number of applications, but one of the coolest classic ideas is to use it for Semantic Search. One can take an encoder model, like the multi-modal UForm, and a web-programming framework, like UCall, and build an image search platform in just 20 lines of Python.
```python import ucall.rich_posix as ucall import uform import usearch
import numpy as np from PIL import Image
server = ucall.Server() model = uform.get_model('unum-cloud/uform-vl-multilingual') index = usearch.Index(ndim=256)
@server def add(label: int, photo: Image.Image): image = model.preprocessimage(photo) vector = model.encodeimage(image).detach().numpy() index.add(label, vector.flatten(), copy=True)
@server def search(query: str) -> np.ndarray: tokens = model.preprocesstext(query) vector = model.encodetext(tokens).detach().numpy() neighbors, _, _ = index.search(vector.flatten(), 3) return neighbors
server.run() ```