Bilibili Intent Filter

Bilibili Intent Filter is a community-built Chrome / Edge extension that filters content cards using natural-language viewing preferences. It reads titles, creators, secondary card text and durations from cards, optionally adds video tags, and sends this limited metadata to a decision model. Updated demo The v2 recording at the top runs for about 2 minutes 10 seconds. It shows homepage scores and blurred cards, scrolling through additional content, and filtering after a page refresh. Jev score badges are distinct from local-rule badges such as ads. This is a workflow demonstration, not an accuracy or latency benchmark. The original Chinese interface is preserved in the video. How filtering works 1. Define profiles such as learning or gaming, with wanted topics, exclusions and blocked keywords. 2. Local keyword and content-type rules run first; cards requiring semantic judgment go to the model. 3. Jev Score assesses relevance and Noul assesses exclusion; code combines these into a 0–1 score. 4. Cards below the threshold are blurred, with temporary reveal on hover. Multiple profiles can be active together. 5. Cards near the viewport are scored on demand, while persistent caching reuses earlier results. New uncached cards or changed settings may still require requests. Model and installation The supplied local variant defaults to typesafe/jev-1.14 through OpenRouter. Settings also support the TypeSafe endpoint, a custom HTTP scoring service and a built-in lexical baseline without a model API. Load the unpacked extension in Chrome or Edge developer mode. Hosted Jev requires your own API key; inference charges depend on the provider. The public source link points to the original author's project, whose default configuration may differ from this demonstration variant. Data and limitations When a remote model is selected, preferences and card metadata are sent to the configured service; inference is not entirely local. API keys are stored locally in the browser. Ordinary settings and preference profiles may sync with the browser account. Rules, caching and visual blurring are handled by the extension. The model does not inspect full video frames, audio or descriptions, so metadata alone cannot guarantee correct filtering. Failed scores leave cards visible, and Bilibili layout changes can affect card detection. This listing is based on source review and the supplied recording. The video passed a complete decode check; we did not reinstall the extension or independently benchmark model quality. Original project: https://github.com/5pr1ng6ridge/bili-intent-filter Local demonstration variant: v0.5.0, commit af4dc8b, with OpenRouter defaults.