What's New
- Complete UI overhaul (#2396)
- LocalDocs improvements (#2396)
- Use nomic-embed-text-v1.5 as local model instead of SBert
- Ship local model with application instead of downloading afterwards
- Store embeddings flat in SQLite DB instead of in hnswlib index
- Do exact KNN search with usearch instead of approximate KNN search with hnswlib
- Markdown support (#2476)
- Support CUDA/Metal device option for embeddings (#2477)
Fixes
- Fix embedding tokenization after #2310 (#2381)
- Fix a crash when loading certain models with "code" in their name (#2382)
- Fix an embedding crash with large chunk sizes after #2310 (#2383)
- Fix inability to load models with non-ASCII path on Windows (#2388)
- CUDA: Do not show non-fatal DLL errors on Windows (#2389)
- LocalDocs fixes (#2396)
- Always use requested number of snippets even if there are better matches in unselected collections
- Check for deleted files on startup
- CUDA: Fix PTX errors with some GPT4All builds (#2421)
- Fix blank device in UI after model switch and improve usage stats (#2409)
- Use CPU instead of CUDA backend when GPU loading fails the first time (ngl=0 is not enough) (#2477)
- Fix crash when sending a message greater than n_ctx tokens after #1970 (#2498)
New Contributors
- @woheller69 made their first contribution in (#2339)
- @patcher9 made their first contribution in (#2386)
- @sunsided made their first contribution in (#2414)
- @johnwparent made their first contribution in (#2319)
- @mcembalest made their first contribution in (#2488)
Full Changelog: v2.8.0...v3.0.0