What's New
Bug Fixes
- Knowledge base indexing no longer aborts on partial failures — when individual embedding batches fail (e.g. due to context length limits), indexing now continues with the remaining chunks and reports partial success instead of restarting the entire run (#579, #581)
- Empty/whitespace chunks are filtered before sending to the embedding API, preventing spurious HTTP 400 errors
- Base URL normalization — entering
https://api.openai.com/v1as the embedding base URL no longer produces 404s from path doubling (/v1/v1/embeddings) - Provider-aware error messages — embedding errors now show the actual provider name ("Ollama" or "OpenAI Compatible") instead of always displaying "Ollama"
Maintenance
- Bumped Homebrew cask to v1.74.5 (#582)
Thanks to @hashmil for the detailed bug report and cross-provider analysis that made these fixes possible.