Unum

GPUs moved to FP8. CPUs didn't. 200K lines of SIMD kernels later, USearch bridges both — native FP8 search & KV caching on NVIDIA's new Vera CPU racks, Intel Xeons, RISC-V, IBM mainframes, Chinese LoongArch, & WebAssembly. 99% recall at half the RAM! unum.cloud/blog/float8

Despite being the most popular vector search engine in the Python ecosystem, most users of Meta's FAISS library don't seem to know its strengths or weaknesses. Since FAISS is the most common comparison target for (my) @unum_cloud USearch engine, I do a fair share of benchmarking against it — and a few things keep coming up that are worth sharing. FAISS ships several scalar quantization modes that almost nobody uses. If you are dealing with `f16` or `bf16` vectors, you pass `SQfp16` / `SQbf16` as qualifier strings to `index_factory`. For native `u8` or `i8` integer inputs there's `SQ8_direct` and `SQ8_direct_signed` — the "direct" variants store the byte as-is, without per-dimension rescaling, which is exactly what you want when the data is already in range. For the most extreme case — binary vectors — you shouldn't use the regular `HNSW` class. FAISS has a separate `BHNSW` instantiation. It's less configurable: you can't set `efConstruction` or `efSearch` via the C API, they're hard-coded at 40 and 16. But it's dramatically faster than shoving binary data through the float path, and at those defaults it still hits reasonable recall. The bigger problem is that FAISS is tuned for fairly small-scale use. As indexes grow, construction and search speed degrade super-linearly. On a 192-core machine I couldn't finish indexing fairly modest 100M datasets — both from BigANN (SIFT 100M `u8` with L2) and from the binary-fingerprint datasets I've published for computational chemistry and biology (PubChem MACCS, 168-bit, Hamming distance). The same hyperparameters that run in minutes at 10M sit for hours at 100M without finishing. It's worth noting that FAISS leans heavily on hybrid schemes with learnable quantization — IVFPQ and friends. I stand by the opinion that those approaches aren't robust: they give short-term gains on small datasets and backfire at real scale. That's a topic for a proper paper someday. For now, the advice is simple: dig through the docs for forgotten features, and run your own benchmarks to understand the strengths and limits of the tools you're using. You'll probably find knobs nobody talks about. PS: I was working on some hybrid CPU+GPU workloads on @nebiusai — check out the numbers I got on their new Intel Xeon 6 Granite Rapids + Nvidia Blackwell instances 🔥 github.com/ashvardanian/R…

USearch v2.25 is out — adding Float8 support for upcoming NVIDIA Vera CPUs and 1000+ new kernels across x86, Arm, RISC-V, PowerPC, LoongArch & relaxed WASM SIMD 🥳 github.com/unum-cloud/USe…

Open source USearch library jumpstarts ScyllaDB vector search thenewstack.io/open-source-us…

New milestone for Unum in the Rust ecosystem! USearch now outranks Meta’s FAISS on Crates.io by 4× in recent downloads — and the underlying SimSIMD vector math library by 30× 🎉

VC work — on both the investing & founding side — pulled some time away from R&D these last few weeks. But as @unum_cloud turns 10, it feels like the right moment to finally put up a proper page for one of the world’s most widely used search engines 🥳 unum.cloud/usearch

USearch is getting close to earning the title of "SQLite of Search" 🥳 After C++11, C99, Python, Rust, JS, Java, Obj-C, Swift, C#, Go, Wolfram, Kotlin, and Clojure — @ZigLang becomes the 14th language with @unum_cloud USearch binding ⚡️ Check out Adib's repo for details 👇

Genome sequencing is advancing faster than Moore’s Law. Great for personalized medicine — but we can’t just keep adding compute. New algorithms & hardware-friendly libs are needed! Our story with @unum_cloud & @nebiusai on porting StringZilla to GPUs: nebius.com/customer-stori…

Powered by our MLOps-tailored AI Cloud, @unum_cloud optimized an open-source string processing library StringZilla with hardware-specific kernels to efficiently leverage GPU parallelism at the software layer. Read the full story: nebius.com/customer-stori… #StringZilla #Strings

The 26-year-old Apache Lucene is my worst nightmare. It’s arguably one of the world’s top 3 search engines — powering Elastic, Solr, MongoDB Atlas, AWS OpenSearch, and Azure Cognitive Search. Alongside Meta’s FAISS and (my) @Unum_Cloud’s USearch — the younger successors. So yes, I’m pretty biased! To give JVM projects a sensible upgrade path to something 10–100× faster, I wired both Lucene and USearch into Apache Spark — the engine at the heart of Databricks — and ran them on a 192-core AWS box. Part fiasco, part success: I couldn’t find a Lucene knob combo that delivered remotely adequate performance. Might be a skill issue, though I also poked several much more experienced Lucene users & contributors. For USearch, Spark turned out to be the perfect test-bed: Maven-free Gradle builds that pull “fat JARs” from GitHub and auto-link to precompiled, SIMD-accelerated binaries for the current OS and hardware platform. Exactly the kind of migration UX I wish I’d had years ago. After the first USearch releases, some Elastic/Solr users reported rough edges during migration. Now I see why — and those edges are finally getting filed down 🤗 I shared all of this yesterday at the PyTorch London meetup hosted at @Databricks — a great crowd to share my fiascos! Huge thanks to the organizers and community for the invitation 🙏 Slides: attached 🧵

Although most of our open-source libraries focus on in-memory processing, truly large-scale processing typically begins in external memory. Here's a glimpse from 2022 into what has been in the works for quite some time and will hopefully be available to everyone 🔜

USearch adoption is snowballing — well beyond ClickHouse, DuckDB, YugaByte, TiDB & ScyllaDB open-source gurus and the vector search use-case to the contents of slide 41 of this 2023 lecture and further! So many ways to exploit just a few thousand lines of C++ templates ⚛️

USearch everywhere 🎉 @Unum_Cloud 🤝 @Yugabyte

Higher recall + Python, JS, & Swift fixes🎉 Entirely community-driven - I did nothing 🤗 Upgrade to USearch v2.17.8 today 😉 github.com/unum-cloud/use…

For everyone using @unum_cloud USearch For <1B vectors per index you should aim for >100K insert & search queries/second with reasonable accuracy on 1 node 100M vecs ~ 20 mins I’ve met several AI labs, using it and not realizing its 10x faster 😉

One more DBMS project leveraging @unum_cloud USearch for Vector Search 🥳 github.com/unum-cloud/use…

20 million Python downloads across SimSIMD, StringZilla, and @unum_cloud USearch 🎉 If you’re chasing performance, upgrading to Ubuntu 24.04 LTS is worth it — modern Linux kernels, compilers and improved SIMD intrinsics support make a huge difference! github.com/ashvardanian