feat: Image search example

[Green-Sky]

I used the usearch vector database library, to build a image embedding database and then to search in it by similarity using a text string.

build the database:

$ bin/image-search-build -m ../models/ggml-model-f16.bin ../tests/
clip_model_load: loading model from '../models/ggml-model-f16.bin' - please wait....................................................clip_model_load: model size =   288.93 MB / num tensors = 397
clip_model_load: model loadded
main: starting base dir scan of '../tests/'
main: found image file '../tests/white.jpg'
main: found image file '../tests/red_apple.jpg'

it generates an images.usearch and images.path file.

search by similarity to string:

$ bin/image-search -m ../models/ggml-model-f16.bin "a red apple"
clip_model_load: loading model from '../models/ggml-model-f16.bin' - please wait....................................................clip_model_load: model size =   288.93 MB / num tensors = 397
clip_model_load: model loadded
search results:
similarity path
  0.655244 /home/green/workspace/clip.cpp/tests/red_apple.jpg
  0.773273 /home/green/workspace/clip.cpp/tests/white.jpg

It is a very rough implementation. It requires c++17 and std::filesystem for iterating the filesystem.

A similarity search via an image would be nice too.

basically we could copy all the features from here: https://github.com/yurijmikhalevich/rclip

TODO:

• add README.md to example
• make it compile conditionally (bc c++17 and deps)
• solve the usearch memory leak (tracking Bug: index.hpp leaks memory unum-cloud/usearch#138)
• tabs to spaces
• use fetch content instead of committing deps

[monatis]

I like the idea to provide an image search example with tiny vector search implementation directly in C++, but I highlighted some points.

Also, let's not use tab characters for indents please.

general code clean up. but the rest of the repo is still in a similar state 😝

Things may be messy at early stages to quickly introduce new features, but I'm trying to tidy up the code and README as we make progress in the project. You can always suggest better structures for the code and the overall project.

[Green-Sky]

Also, let's not use tab characters for indents please.

Ohhhhh yea. sorry. forgot to reverse this 😆

[Green-Sky]

Things may be messy at early stages to quickly introduce new features, but I'm trying to tidy up the code and README as we make progress in the project. You can always suggest better structures for the code and the overall project.

yea, I also intentionally maked this PR as draft for those reasons. :)

[monatis]

yea, I also intentionally maked this PR as draft for those reasons. :)

Great, thank you. Your work really appreciated. And to be honest, I was not aware of Usearch, but instead considering the original hnswlib library for the search example. But Usearch looks like a better candidate.

[Green-Sky]

forcepushed with FetchContent and spaces instead of tabs.

[monatis]

Do you need an extra hand in this?

[Green-Sky]

Was doing something completely different last week :)
Will try to get this into a mergeable state this week.

[Green-Sky]

just rebased and changed some stuff you committed on main

issues that I found:

1. you added cmake options for conditionally disabling tests and examples, i negated them and initialized them to standalone (more convenient)
2. you require c++20 (!) should only be 11 i think. (and 17 for fs in image-search until its changed to something else)
3. you set gcc/clang specific warning and O3 optimization level indiscriminately

   clip.cpp/CMakeLists.txt

   Line 136 in 1f140e0

   add_compile_options(-Wno-format -O3)

   (you should not, you default to Release

   clip.cpp/CMakeLists.txt

   Line 7 in 1f140e0

   set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)

   , which sets O3)

[monatis]

Yes CMakeLists.txt still requires some testing and optimizations (also for other arches like different arm versions) so it might be another PR when I take down my Raspberry PIs from the shelf. C++20 standard was just a guard but we can downgrade it now or after some tests. I can refactor the directory walk with the function from common-clip that does not require C++17. Then everything will be ok with C++11.

I'm ok with it as is, and then we can continue to further improve it. When #30 is merged, we can also the use batch inference when indexing images.

[Green-Sky]

Yes CMakeLists.txt still requires some testing and optimizations (also for other arches like different arm versions) so it might be another PR when I take down my Raspberry PIs from the shelf. C++20 standard was just a guard but we can downgrade it now or after some tests. I can refactor the directory walk with the function from common-clip that does not require C++17. Then everything will be ok with C++11.

yea, there have also been changes upstream (especially llama) , where you obviously copied based the code on.

I'm ok with it as is, and then we can continue to further improve it. When #30 is merged, we can also the use batch inference when indexing images.

Yea. Let me run some small tests and then i will make it ready for review.

[monatis]

yea, there have also been changes upstream (especially llama) , where you obviously copied based the code on.

Exactly. The entire ggml-based projects are moving very fast, so I believe we don't need to target 100% production-grade multiplatform-ability. The speed of iterations and innovations is prioritized currently.

Yea. Let me run some small tests and then i will make it ready for review.

Great. Thank you.

[monatis]

USearch reports distance values instead of similarities, so I made changes to emphasize it. LGTM, merging. TODOs can be fixed later.

Thanks for your contributions.