Local SQLite FTS5 search

The user-facing part.

Find saved items through library search. Clip search goes inside videos to return timestamped, deep-linked excerpts, and evidence cards collect material your AI can quote and cite. Search no longer needs every word: it skips common words and ranks sources with every word and exact titles first. Choose Exact words when you need every word to match. Optional semantic search finds passages with similar meaning; turn it on in Settings, and it downloads its model once (about 330 MB) and runs on your computer.

Run the loop.

1. Open the Uoink desktop dashboard.
2. Click the search bar at the top of the interface.
3. Type your search query, such as a topic, speaker name, or tool.
4. View the matching corpus cards with highlighted search snippets.
5. Click any card to open the markdown file.

How Uoink handles it.

The stdio tools search_library, get_library_item, search_clips, and get_evidence_card expose saved source material to MCP clients. Keyword search runs on the local SQLite full-text index with BM25 ranking. Optional semantic search keeps an EmbeddingGemma vector index on your computer. search_library, search_uoinks, and search_clips take optional match and mode parameters, and each result cites a passage only when that passage contains the match.

Agent-readable surface.

Full server metadata lives at /mcp and /mcp/manifest.json.

Run it on your next source.

Search across your entire captured library instantly without cloud latency.