USearch v3: Text Indexes, Faster Views, Cleaner API and Docker

[ashvardanian]

The v3 release of USearch is bringing several new features, and provides an excellent opportunity to refactor existing ones!

• Uniform naming across bindings.
• Making tqdm and ucall dependencies optional.
• Co-locating vectors and graph data on disk.
• Better compaction before serialization to make searches over view faster.
• TextIndex, TF-IDF, and Levenshtein distances from StringZilla.
• Docker image with a UCall-based REST API.
• Tooling for sampling/pooling a hybrid embedding from long documents embedded in chunks.
• Support batch-capable distance metrics and masked loads from SimSIMD.

Candidates:

• External retrievers and node-level serialization. @Ngalstyan4
• Batch-add and batch-search interfaces for C, Rust. @var77

[var77]

Thank you for your work! It will also be great to have a batch add functionality in Rust bindings.

[monatis]

Thanks for the awesome work!

• Docker image with a UCall

Will this support multiple indexes / collections? Like endpoints for creating a collection (vector index), listing available collections, searching in the given collection etc. all in a single Docker container.

[ashvardanian]

@monatis, yes, that’s the plan 🤗

[monatis]

Super news. It's already great for embedded use but a dockerized service mode can cater for new use cases.

[ashvardanian]

I've released SimSIMD v2, which will be included in USearch v3. It brings several performance and versatility improvements and may soon feature Intel AMX support and more bitwise metrics.

[philippemnoel]

Thank you for your work! Do you have an estimate on when USearch v3 will be ready?

[ashvardanian]

Thank you, @philippemnoel! I plan to release it before the end of October.

Here is the progress so far:

• A better compaction plan for on-disk views. cc @al13n321.
• Masked-loads support in SimSIMD shipped this weekend.
• Levenstein distance present in StringZilla will be released with the next major release, potentially tomorrow.

Once we fix the build issues in UCall CI, all the pieces will fall into place, and the release will be easier to prepare. Feel free to help there 🤗

[aehlke]

Would be amazing to prioritize langchainJS integration. I can run it on web and mobile and desktop easily with that. You'll get attention more quickly from their many users.

[ashvardanian]

@aehlke, I believe USearch is already available in LangChainJS.

[aehlke]

Ah I meant for this new text search functionality - maybe there's nothing to be added on their side. Looking forward to this PR, thanks for the work.

[ashvardanian]

🎉 This PR is included in version 2.8.3 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀

[monatis]

Hi @ashvardanian it seems that this pr is merged without V3 changes. I just wanted to ask if you are still planning to release them and if you have a new ETA?

I was thinking of using USearch in an upcoming product but a stable file format might be important.

Not necessarily related to V3, but are you planning to support incrementally adding vectors in memmaped files, or what is the recommended procedure to add new vectors after the initial upload?

[aehlke]

Is there an update on text index and related updates originally planned? Thanks

[ashvardanian]

Hello @monatis and @aehlke,

I hope you're both doing well. The past few weeks have been incredibly busy as we wrapped up our year-long projects. You might have come across some of these initiatives:

• Publishing 28 Billion Molecule Embeddings with AWS
• Scaling USearch Vector Search Engine with Gcore on Intel's Sapphire Rapids CPUs

Collaborating with large organizations has its challenges, particularly in terms of pace. This has required us to adjust our timelines slightly. Nonetheless, progress on v3 is steady, along with other updates and integrations, including contributions to SciPy and a potential integration with Sklearn.

Beating existing solutions in vector and text search, clustering, dimension reduction, and external memory access is doable. But achieving that in one package under 10,000 lines of maintainable code compatible with every OS and hardware architecture is very tricky. I want to make sure the design persists for years, so I'm not rushing to make sure we get it right.

Thank you for your continued support and patience! 🤗

[aehlke]

I appreciate the context, best of luck!