Changelog
v0.1.0 (Unreleased)
Added
- Int8 quantized model: default model is now 7.5MB (down from 29MB f32)
- Auto-quantization: f32/f16 models from HuggingFace are automatically quantized to int8 on download
- Token-level coloring: search results highlight individual tokens with a relevance gradient (gray → red)
- WASM search demo: live in-browser semantic search at sgrep.sh, powered by the same model compiled to WebAssembly
- Heading-chunked docs search: ⌘K search links directly to the relevant section heading
- Context lines:
-C,-B,-Aflags for showing surrounding lines (grep-compatible) - Caching: sqlite-vec index for instant repeated searches (
--index,--reindex,--no-cache) - Hybrid search: combine BM25 keyword matching with semantic similarity (
--hybrid,--alpha) - Zero-shot classification: label text chunks against arbitrary categories (
-e) - Code search model:
sgrep-code-v1fine-tuned for code search - Mobile docs: responsive layout with hamburger menu and semantic search
- curl installer:
curl -fsSL sgrep.sh/install.sh | sh
Changed
- Model downloads from sgrep.sh instead of HuggingFace (faster, smaller)
- Stale f32 model caches are auto-detected and re-downloaded as int8
v0.0.1
Initial release. Pure Rust semantic grep with Model2Vec embeddings.
Was this page helpful?