Devotd Github
NewRAG & retrieval

anush008/fastembed-rs v7.6.1 — fix for the finite-value cosine similarity bug

anush008/fastembed-rs v7.6.1 is a bug-fix release that ensures the cosine similarity scores of finite vectors stay finite.

Repositoryanush008/fastembed-rs ↗

anush008/fastembed-rs v7.6.1 is a bug-fix release that ensures the cosine similarity scores of finite vectors stay finite.

GitHub description: Rust library for generating vector embeddings and reranking locally!

stars
★ 1K
Forks
145
Language
Rust
License
Apache-2.0
Latest release
v7.1.1
Read original source ↗View on GitHub ↗

anush008/fastembed-rs is a Rust library that generates vector embeddings and reranking locally. v7.6.1 was released in commit d32c197 on 2026-11-06, and the contents of this version consist of a single bug fix in the similarity module. Even when all input vectors are finite values, no NaN or infinity gets mixed into the cosine similarity score any more — finite inputs now produce finite results. This fix is contained in commit 678001a, which the release message points to, and it addresses issue #350.

Cosine similarity, used when comparing embedding vectors, is a metric that looks only at the direction of the two vectors. The problematic point is when the score is not finite. If such values are put into a sort criterion or a threshold comparison, the ordering of the results can become abnormally mixed up. The library handles this calculation in the similarity module, and this patch corrected that boundary condition. This repository is a library aimed at producing embeddings and reranking locally, and Cargo.toml declares edition 2021 and a minimum Rust version of 1.88.

That said, the patch attached to this commit is only a single version line in Cargo.toml (from 7.6.0 to 7.6.1). The actual code fix is in the separate commit 678001a, and this release commit itself is marked with [skip ci] so that it skips the automation pipeline. The collected records contain no performance or accuracy figures for this fix, and no license and star count listings. So this version is better regarded not as a release that changes embedding quality itself, but as a maintenance patch that fixes the boundary condition of the similarity calculation. If NaN has ever popped up in code that uses similarity, it is worth upgrading to 7.6.1; if you have had no such experience, there are not many new items to check.