turbovec
Written in Rust with Python bindings, turbovec implements Google Research's TurboQuant algorithm to compress vector embeddings into 2-bit or 4-bit representations without a separate training step. The library targets local retrieval-augmented generation pipelines where developers need online vector insertion, query-time candidate filtering, and incremental on-disk updates without running external infrastructure. It provides integrations intended to swap into in-memory store interfaces across LangChain, LlamaIndex, Haystack, and Agno.
Adopting turbovec requires float32 input arrays of shape (n, dim), rejecting other numeric types at ingest. Its release 1.0 commit standardizes on-disk format version 7, breaking compatibility with older .tv and .tvim files. Hardware-specific execution relies on hand-written SIMD paths—AVX-512 VNNI and AVX2 on x86, NEON on ARM—with scalar fallbacks on older CPUs, while targeting an x86-64-v2 compiler baseline. The project is licensed under the MIT License.
nobuzz
The nobuzz repository supplies /debuzz, a custom skill for Claude Code that rewrites verbose or theatrical assistant explanations into direct technical summaries. Aimed at engineers who want concise bug analyses without rhetorical flourishes, the utility accepts optional modes (colleague, manager, or director) to tune output length and detail for different audiences.
Operation depends on external tooling: users must install Google's Antigravity CLI (agy), authenticate via Google Sign-In, and manually place the skill directory inside ~/.claude/skills. The tool functions by writing the prior assistant message to a temporary local file and passing its absolute path to agy in headless mode (-p) to produce a verbatim Gemini rewrite, falling back to a Claude self-rewrite only if the external call fails. The repository is released under the MIT License.