Pay an LLM to Find the Buried Ads in Bilibili Videos

What it does

VideoAdGuard is a browser extension for Bilibili that analyzes video subtitles with a large language model to detect embedded or spoken sponsor advertisements. When it identifies a segment, it overlays a skip button on the player and highlights the ad interval on the progress bar. If a video lacks subtitles, the extension can download the audio stream and transcribe it via Groq’s Whisper API before running the same text analysis.

The interesting bit

The project treats LLM inference as a utility bill you can shop around: it supports OpenAI, DeepSeek, Zhipu, local Ollama, and roughly a dozen other providers through a unified adapter. It also runs a Cloudflare Worker-backed cloud cache so the community can share detection results, falling back to local 24-hour caching when the service is down or the crowd hasn’t seen the video yet.

Key highlights

• Supports both subtitle-based and audio-based detection pipelines for videos with or without captions.
• Plugs into a wide roster of LLM backends, including local deployments via Ollama or LM Studio.
• Caches results locally for 24 hours and queries a community cloud cache (Cloudflare KV) to avoid redundant API calls.
• Lets users manually adjust ad boundaries on the timeline and submit corrections back to the shared cache.
• Includes an auto-skip mode and a per-UP whitelist for channels you trust.

Caveats

• Requires your own LLM API keys; detection quality scales directly with model capability, and token costs are your problem.
• Audio recognition depends on Groq’s Whisper API, which the README notes requires “special network conditions” or a reverse proxy.
• The Tampermonkey version is explicitly deprecated and no longer maintained.

Verdict

Worth a look if you watch enough Bilibili to notice the sponsored segments but not enough to memorize their timestamps. Skip it if you object to feeding video subtitles to third-party AI endpoints—or to paying for the privilege.