nodecode-by2kb

By2kb: finds learning videos by topic and turns a Bilibili or YouTube video into a transcript, an abstract and study notes in your own knowledge base. A Nodecode rewrite of Charlesmpc/by2kb by Charlesmpc.

The work is done by by2kb, a Python CLI: it fetches the media, transcribes it with a local Whisper model (or an optional cloud one), and writes three durable Markdown files. The cell runs that program and reads its JSON; the summaries are written by the session's own model, so the CLI never needs an API key.

Get it

Tell the agent "add the by2kb cell", or run nodecode add nc://by2kb.

Set it up

The cell drives the by2kb program, so that has to be there first:

pipx install "by2kb[asr-whisper,youtube]"     # or: uv tool install --python 3.12 "by2kb[asr-whisper,youtube]"
by2kb init --preset agent-local --library-root ~/knowledge
by2kb models install                          # the local Whisper model, once

It needs Python 3.12+, ffmpeg and ffprobe on PATH. The program is found on PATH; if it lives somewhere else, give its path in the cell's by2kb setting (/settings), and a cell that cannot find it refuses to start and says so. The knowledge base is the CLI's own: ~/.by2kb holds its config, and the files land in the library root you chose.

Use it

Ask for learning material on a topic, and send a video link when you have one. The model gets six tools:

Tool	What it answers
`by2kb_search`	Numbered recommendations for a topic -- title, link and why it matched -- from Bilibili and YouTube. Metadata and cached captions only: nothing is downloaded yet.
`by2kb_select`	Takes the numbers the person replies with (`1`, `1,3`) from the list this conversation was last shown, and starts transcribing and filing those videos.
`by2kb_video`	One video: a Bilibili/YouTube URL, or the path to a local audio/video file. Answers a job.
`by2kb_next`	The job's next staged summary operation: the prompts to answer, its size and time limit, and the operation id.
`by2kb_submit`	Hands one operation its whole answer -- JSON when it asks for JSON, Markdown otherwise.
`by2kb_status`	Where a job stands, and the files it has written so far.

The summaries arrive in stages: a video answers a job, by2kb_next asks for one bounded operation at a time, and by2kb_submit takes the answer, until the job reports the three files -- transcript, short abstract, long-form notes. The abstract is shown in the conversation; the paths are the knowledge base's.

You can look yourself:

/by2kb TiDB vector search      numbered recommendations
/by2kb 1,3                     file those two videos
/by2kb https://youtu.be/...   one video
/by2kb status JOB             where it stands

And from eval: (by2kb:search "TiDB vector search"), (by2kb:video "https://..."), (by2kb:next "JOB"), (by2kb:submit "JOB" "OP" "..."), (by2kb:status "JOB"). (help :by2kb) has the same guidance the model is given.

Discovery is deliberately conservative: a named subject has to appear in a candidate's title or description, and a thin result set is not padded. A recommendation is an excerpt, not the video -- the model is told not to claim it read a full video from one.

Settings

Setting	Default	What it is
`by2kb`	`by2kb`	The program's name on `PATH`, or its path.
`provider`	`nodecode`	The provider name recorded in a summary job's provenance.
`model`	`nodecode`	The model name recorded in a summary job's provenance.
`scope`	empty	An opaque scope a conversation shares, so a number answers its own list. Empty derives one from the session.
`timeout_seconds`	3600	One run's ceiling; transcribing a long video takes minutes.

What is in it

File	What it is
`cell.lisp`	The six verbs, the argv each builds, the JSON each reads, `/by2kb`, and the standing paragraph.
`package.lisp`	The `by2kb` package: the verbs and the condition every refusal signals.
`test/`	Offline tests over a stubbed program: the argv each verb asks for, the recommendation list, the selection, one video's three files, a staged operation read page by page, the submitted answer and the refusals.

The Hermes plugin intercepted a gateway message with a video URL, a by2kb TOPIC prefix or a bare number; Nodecode has no such seam, so here the model calls the tools and /by2kb gives the person the same routes. NOTICE says what else was left out.

nc://by2kb