Capture becomes more valuable when older material can return at the right moment. Drop first added direct Inbox search, then introduced a separate local embedding model, a backfill pipeline, Related Drops, and a semantic search prototype.
The retrieval design keeps direct evidence in front. Exact keyword matches rank above semantic-only results, and result rows explain whether the match came from source text, recognized image text, or a saved transformation.
The semantic layer can be disabled without affecting ordinary search. Its embeddings remain local and can be cleared or rebuilt independently of the Drops themselves.
The capability remains experimental. The production embedding path is English-oriented, semantic quality depends on the input, and active search still evaluates in-memory text rather than a dedicated full-text index.
← Back to Drop