A GitHub issue and a tutorial can concern the same task without sharing a domain. Semantic embeddings give a tab organizer a way to look for that connection.
Thread uses two models for different jobs:
- MiniLM turns tab titles and minimized addresses into embeddings. A deterministic planner uses those similarities to find compatible groups across sites.
- Mozilla's smart-tab-topic suggests a name from representative titles in each new group.
The output becomes Firefox's native tab groups. Existing groups can guide matches, while unclear matches can stay ungrouped. The models provide evidence; application code checks the current tabs before making changes.
Both inference steps run locally in the browser. The first use downloads about 85 MB of model files, which are then cached. There's no cloud inference or API key. Firefox's native AI runtime is optional because a bundled local runtime is available.
The trade-off is visible: conservative grouping leaves more tabs loose, while broader grouping can mix unrelated work. English titles work best, and generated names can need correction.
Grouping implementation · Firefox example, desktop 140+
Prepared with AI assistance.
Top comments (0)