managing-embeddings

To generate semantic embeddings for notes, you can call the Gemini embedding API for each note, preprocess the text to remove HTML, and store the resulting vectors in a local file-based cache.

Cosine similarity scoring evaluates the mathematical distance between generated note embeddings, enabling semantic matching across decks to retrieve content with similar meaning rather than exact keyword matches.

Yes, you need a Node TypeScript runtime to execute the embedding generation logic, manage the file-based cache directories, and run text preprocessing tasks for the Gemini API integration.

Yes, you can update the embedding cache incrementally by applying simple invalidation rules to process only new or modified notes, avoiding the need to regenerate embeddings for the entire deck.

Text preprocessing for semantic search normalizes input content by removing Cloze syntax and HTML tags, and it is required before generating embeddings to ensure the API receives clean text for accurate vector representation.

You can monitor embedding health and coverage for a specific deck by running deck-level analytics, which evaluate cache status and identify missing or invalid semantic vectors across your notes.