v0.6.66b490

Search that scales to millions of pages

Search used to scan every vector in the index on every query. That works for a personal vault but falls over on large libraries. lilbee now builds an approximate vector index automatically once a vault grows past 50k chunks, so search stays fast at millions of pages. Smaller vaults keep exact search, so nothing changes for everyday use. (#321)

Downloaded indexes are also plug-and-play now. An index only works with the embedder that built it, and previously lilbee would tell you to rebuild one you had just downloaded. Now it offers to switch to the index's own embedder instead, downloading it if needed, from the TUI, the CLI (lilbee use-embedder), or the Obsidian plugin.

Export and import your library as a dataset

You can now get clean per-page text back out of a lilbee index as a {source, page, text} dataset (parquet or jsonl), and import one back under whatever embedder is active. This makes it possible to share a library, publish its text, or move between embedding models without rebuilding from the original documents. Available everywhere: lilbee export / lilbee import, /export and /import in the TUI, MCP tools, and the HTTP API with live progress. (#322)

Long-term memory for chat

Chat can now remember things across sessions: facts about you, recalled when relevant, and standing preferences for how you want answers, applied to every chat. Off by default, and available from the TUI (/remember, /memories), the lilbee memory CLI, MCP (private per agent), the REST API, and Python. Memories survive index rebuilds and are only removed by a factory reset. (#319)

Fixes and maintenance

• Fixed /add path completion on Windows stripping the directory from backslash paths, which prevented files from being indexed (#316)
• Bumped pyjwt and aiohttp to patched versions for upstream security advisories (#317)
• Updated GitHub Actions dependencies (#315)

Full Changelog: v0.6.66b489...v0.6.66b490