When indexing a small number of vectors I am getting this error when specifying cosine_similarity (euclidean works fine for instance):

Thanks for the very nice library! I'm interested in using hora for doing nearest neighbor finding in single-cell genomics. The data of interest consist of very high dimensional points (D = 30,000), but for most points, most dimensions have value 0. Therefore, I'd like to avoid (it's not really feasible) to densify the elements before indexing them. Is there some way to provide a custom implementation of the relevant distance metrics for the indexed type such that I don't have to actually insert a dense representation of the points into the index?

as title

This line assumes, that dot function returns a simple dot production and negates it to make it work like a distance (smaller = closer) https://github.com/hora-search/hora/blob/a6759f8ae950a10d833735b9bca82d80435e60aa/src/core/metrics.rs#L56

But in reality, dot function is already negated in the SIMDOptmized inplementation: https://github.com/hora-search/hora/blob/ce10e42bacbf29163fcb9a973e1a3f9dc528effb/src/core/simd_metrics.rs#L32

Which leads to double inverting and incorrect use of Dot Production

Can this library add/remove/modify nodes to builded graph?

Hello! Thanks for your wonderful library. Please tell me if supported extensible index. Can I adding elements to an already built index?

Hey guys,

This isn't an issue. You can close or delete it whenever you wish. I just wanted to share a fun fact.

I discovered "hora" at the top of a "most trending projects" lists and it attracted my attention as "hora" ("хора" in Cyrillic) means "people". I thought, is it possible that someone will name his trending lib after a Bulgarian word? Maybe there was a hidden meaning... I opened and scrolled the repo - what can I see - a list of faces/people. I was right. Someone named the library after a Bulgarian word... only to realise a few moments later that it's named after a Japanese word.

What a funny collision.

Have a great week!

For the Node version, is it possible to run comparisons on arrays of objects?

For example, for using it with map data, the query would be a tuple of a map point coordinates: [latitude, longitude].

The material to compare against would be an array of objects similar to:

[ { objectId: 123, someOtherObjectData: string, coordinates: [latitude, longitude] }, { next object }, { next object} ... ]

If so, would it be possible to provide a few pointers or a basic example?

Thanks much!

I am finding a product which can replace Elasticsearch.

So, I tried installing using the pip install horapy, but getting this error:

Hi.

With the latest hora, I get a panic when using hora::core::metrics::Metric::CosineSimilarity.

I think this is an error caused by partial_cmp. partial_cmp expect it to be None when the given value cannot be ordered.

Are there any plans to support Android in the near future?

I'm not sure if this is my problem, (possible). I am testing the index types all work, (SSG had extensive index times > 24 hours so I aborted that). IVFPQIndex

I am doing this with all index types, no issues except this case.

FYI: print(i,d) 15746524 [248, 225, 188, 223, 199, 174, 144, 146]

Error:

Other ANN libraries return the index array of the closest vectors and optionally the distance metric in a second array. Is that possible or in future plans? Thanks.

The current index add method adds a single vector. Is it possible to add 2D numpy array of vectors and use the array row position as the index.

version does is not available.

https://www.python.org/dev/peps/pep-0396/

import horapy

dir(horapy) ['BruteForceIndex', 'HNSWIndex', 'HoraANNIndex', 'HoraBruteForceIndexStr', 'HoraBruteForceIndexUsize', 'HoraHNSWIndexStr', 'HoraHNSWIndexUsize', 'HoraIVFPQIndexStr', 'HoraIVFPQIndexUsize', 'HoraPQIndexStr', 'HoraPQIndexUsize', 'HoraSSGIndexStr', 'HoraSSGIndexUsize', 'IVFPQIndex', 'PQIndex', 'SSGIndex', 'builtins', 'cached', 'doc', 'file', 'loader', 'name', 'package', 'path', 'spec', 'horapy', 'numpy'] horapy.version Traceback (most recent call last): File "", line 1, in AttributeError: module 'horapy' has no attribute 'version'

Hey, I just saw this on Reddit and I am very excited to try this out on my computer vision datasets.

I just created an ANN search data structure somewhat recently called HGG (https://github.com/rust-cv/hgg), and I would love to be able to add it to your collection of nearest neighbor searches (and especially over at https://github.com/hora-search/ann-benchmarks). It is based on HNSW, and I am currently using it for computer vision purposes. Let me know if you need some help integrating it.