Changes
- semantic message search now ranks across the full compatible local vector set instead of only the newest candidate window. (#36) Thanks @GaosCode.
- hybrid message search now fuses FTS with local semantic vectors while avoiding embedding-provider calls when no local vectors exist. (#37) Thanks @GaosCode.
- local embedding providers now support OpenAI-compatible endpoints, Ollama, and llama.cpp, and
doctorcan probe the configured provider before you queue vectors embednow drains the queued embedding backlog in bounded batches, requeues safely on provider throttling, and drops stale stored vectors when messages no longer have embeddable content- Git snapshot publishing can now opt in to backing up generated embedding vectors with
--with-embeddingswhile still keeping embedding queue state local. - Git-backed snapshot imports are now much faster on large archives by using import-only SQLite pragmas and bulk-load FTS5 settings during search index rebuilds
messagesandmentionsnow use composite read-path indexes so larger archives spend less time sorting/filtering common guild, channel, and author queries
Fixes
- normalized message text is now sanitized before it reaches SQLite and FTS5, repairing malformed UTF-8 and stripping invisible/control-character noise that can poison search content
- Git-backed snapshots now keep embedding queue state and generated vectors local to each archive, so subscribers no longer inherit misleading embedding backlog metadata. (#38) Thanks @GaosCode.
Docs
- docs now cover semantic and hybrid search setup, embedding privacy, Git snapshot behavior, and local vector rebuilds. (#39) Thanks @GaosCode.
Tests
- Git embedding snapshot export/import now has CLI, share-package, and Docker E2E coverage.
- total Go test coverage now reaches the 85% line.