Skip to main content
Meaning Search helps when you remember an idea but not its wording. Use exact text search for names, codes, and identifiers. Similarity is a useful lead, not proof of relevance.

​
Connect and enable

  1. Follow Local AI setup to install Ollama, download an embedding model, and connect it in Intelligence → Models.
  2. Choose an installed embedding-capable model under Meaning Search and select Enable.
  3. Check Intelligence → Indexing for progress. Validation and indexing take time; exact search stays available.
  4. Enable the Meaning Search source using the clipboard search source controls.
Eligible text comes from search projections of supported representations and available text artifacts. This is not general image similarity or a search of every file on your computer. Eligibility depends on captured content, artifacts, and configured sources.

​
Try your first result

Copy this harmless note as one clip:
Confirm it appears in history and becomes indexed. With Meaning Search enabled, try a cafe for our group meeting. The clip may appear despite different wording. Check the original text before using it. Ordering depends on the model, threshold, and other clips. This exercise does not guarantee a particular rank. Compare with an exact query such as Kissa Mori to understand the two search modes.

​
If results are missing

  • Check model enablement and whether the semantic source is selected.
  • Review indexed versus eligible clips, pending work, and failures in Indexing.
  • Confirm the clip has searchable text. OCR availability depends on platform and configuration.
  • Try a simpler query and review minimum similarity: stricter thresholds can omit related results.
  • Download missing models through Ollama and refresh the connection.
  • For connection errors, follow Local AI recovery.

​
Manage the derived index

Use Intelligence → Indexing to index missing content, retry failed jobs, or rebuild. Clearing or rebuilding the semantic index preserves original clips and exact text search. Changing embedding models builds a replacement index. Wait for readiness before comparing retrieval. Disable stops embedding; clearing stored index data is separate. Read the app’s confirmation and diagnostic before choosing an action. Recall adds generated answers with sources. It needs generation; embeddings alone do not write answers.