An AI study buddy for Bilibili and YouTube

What it does

BibiGPT v1 ingests video content—specifically from Bilibili and YouTube—and returns an AI-generated summary. It fetches video content, packages it into a prompt, and sends it to an OpenAI-compatible API through a Vercel Edge function, then streams the response back to the browser as readable notes. The project also offers a browser extension and an alternate hosted instance, though this repository is explicitly the v1 codebase.

The interesting bit

The architecture is aggressively cost-conscious. It uses Upstash Redis for caching and rate limiting to avoid redundant API calls, and the author explicitly recommends using text-curie-001 rather than a larger model for summarization. That focus on keeping inference cheap is unusual for a GPT-wrapping side project.

Key highlights

• Streams summaries through Vercel Edge functions rather than blocking on a full generation
• Caches results in Upstash Redis to avoid paying twice for the same video content
• Ships as a web app with an alternate address and a browser extension
• Bilingual branding positions it as both a “stream-saving artifact” and an AI class representative

Caveats

• This repository is explicitly the v1 branch and supports only Bilibili and YouTube, despite marketing language that lists TikTok, podcasts, and meetings
• Docker support is referenced via an external pull request, leaving its maintenance status unclear
• The README mixes Chinese and English instructions without clear separation, which may slow down contributors

Verdict

Worth a look if you want a reference architecture for cheap, streamed GPT summarization of video content. Skip it if you need a fully maintained, general-purpose media parser—this is a v1 snapshot tied to specific Chinese and Western video platforms.