We tested dozens (maybe >100) of papers/ideas over the last years in vision and multimodal perception and this is one of the rare cases, where everything worked well! Neat idea and paper!
This model, for example, uses 4 register tokens, and combines them with Matryoshka-style losses for training, resulting in super-compact 64-dimensional embeddings, in case anyone is looking for CLIP alternatives: https://huggingface.co/unum-cloud/uform3-image-text-english-...
Even the fact that a popular outlet is covering that a language foundation (for a language used by millions) is hiring 1 engineer with the help of a trillion-dollar company with over 200K employees is ridiculous.
You are right. I don't yet support pluggable storage systems, but you can check the Lantern's fork of USearch. It may have the right capabilities, as they wanted the same level of integration for Postgres.
Our current SQLite extension brings "search APIs" to SQLite but not the index itself. Think of it as a bundle of SIMD-vectorized Cosine/Dot/L2/Hamming/Jaccard... (for vectors) and Levenshtein/Hamming/NW (for text) distances coming from SimSIMD and StringZilla. Unlike USearch, the latter are pure C 99 header-only libraries, in case you need smth similar.
Startups often avoid Bio and Pharma because of how slow the big partners can be. Would be great to see other Pharma companies restructuring in similar way - reviving the sector and attracting startups and new technologies.
I often use it even in plain loops, helps me avoid nesting ( https://github.com/ashvardanian/SimSIMD/blob/3e51bacb1b74a7e... ), but for most programmers it probably only associates with `goto cleanup;` used to destroy variables allocated earlier in the scope in case of an error.