WebAI StudioCatch near-duplicate bug reports before they are filed, even when they share no words with the original.
Deduplication runs while the user is still typing, because every similarity check is a local vector comparison — no server calls per keystroke.
Issues appear here once the index is built.
56 demos, all running on-device.
The production pattern for cheap moderation: embeddings clear the obvious 90% in microseconds, the LLM judges only the borderline.
Fully local RAG: retrieve the relevant passages with embeddings, then answer with the Prompt API — citations included.
Search a help center by meaning, side by side with keyword search — and watch keywords miss what embeddings find.