Try semantic search on your own short passages. Pocket Embed compresses MiniLM’s search vectors to one third of their original size.
512 bytesper float32 vector
86–98%of baseline task nDCG@10*
In your browsertext stays on this device
Up to 32 short passages. First search downloads about 23 MB of model weights plus runtime files. Later queries reuse the passage vectors until you edit the passages.