An offline AI archivist for Bilibili and YouTube

What it does BiliSum is a desktop app and CLI tool that ingests Bilibili, YouTube, and local video files, transcribes them through a choice of ASR engines, and runs the results through an AI pipeline. The output is a set of structured summaries, chapter timelines, mind maps, and a local searchable knowledge base. Notes, indexes, and configuration are stored locally, and the stack can run fully offline using local embeddings and LLMs.

The interesting bit The visual note generator is the detail that elevates it above a typical transcript-and-summary tool. A configurable VLM re-reads the transcript alongside candidate video frames, filters low-quality shots, and weaves images directly into the relevant paragraphs rather than dumping them in an appendix. The whole system is also exposed as an Agent skill, so tools like Codex or Cursor can invoke summarization or transcription on demand.

Key highlights

• Supports Bilibili QR-code login to handle site-specific access and anti-bot measures.
• Produces three note styles: plain text, timestamped illustrations, or VLM-driven layouts with inline frames.
• Knowledge base features cross-video semantic search, keyword retrieval, RAG Q&A, and a tag relationship graph.
• Runs locally by default: supports local Whisper, local embeddings, and local LLMs, with data stored in SQLite.
• Exports to Markdown and Obsidian, with optional packaging of screenshots alongside notes.

Caveats

• Desktop builds are only provided for Windows and macOS; Linux users are limited to Docker or source builds.
• Bilibili’s anti-bot measures can interfere with access; the README suggests relying on the desktop app’s built-in QR login to mitigate this.

Verdict Worth installing if you treat streaming platforms as a learning feed and want an offline, Obsidian-friendly archive. Skip it if you only need a one-off transcript—Whisper alone is lighter.