Skip to content
bili

Discovering

Walk the graph of related videos and creators, breadth first, streaming one record per node.

Every other command answers one question about one object: a video's related list, a creator's catalogue, the comments on a post. discover chains them. From a seed video or creator it follows the object's edges, and from each neighbor it follows theirs, hop by hop, streaming one record per node as the node is reached.

bili discover BV17x411w7KC

A seed is anything bili can resolve to a video or a creator: a BV/av id, a video URL, a b23.tv short link, or a space.bilibili.com creator URL.

The graph

There are two kinds of node, videos and creators, and three edges between them:

Edge From → to What it follows
related video → video the videos bilibili recommends alongside one
uploader video → creator the creator who posted a video
uploads creator → video the videos in a creator's catalogue

You rarely name edges one at a time. --follow takes a preset:

Preset Expands to Walk shape
content (default) related + uploads stays among videos: a video's recommendations, a creator's uploads
creators uploader + uploads bounces between videos and the people who make them
all every edge the whole neighborhood 
bili discover BV17x411w7KC                       # content (the default)
bili discover BV17x411w7KC --follow creators     # video → uploader → their uploads
bili discover BV17x411w7KC --follow all --depth 2

--follow also takes a single edge name, or a comma-separated mix of presets and edges, so you can be exact:

bili discover BV17x411w7KC --follow uploader        # just hop to the creator
bili discover BV17x411w7KC --follow related,uploader

Bounding the walk

Three independent limits keep a walk finite:

  • --depth is how many hops to follow (default 1; 0 emits only the seeds).
  • --fanout caps neighbors per edge (default 25).
  • -n caps the total nodes streamed (default 500), so an unbounded discover BV… always terminates instead of spidering forever.
bili discover BV17x411w7KC --depth 3 --fanout 10 -n 200

Reading the output

Each row is a node tagged with how it was reached: how deep, by which edge, the object and its owner. The full typed record rides along for -o json/-o jsonl, and -o url prints one link per node:

bili discover BV17x411w7KC                 # the readable table
bili discover BV17x411w7KC -o jsonl        # one lossless object per line
bili discover BV17x411w7KC -o url          # one URL per node, to pipe onward

Seeds can come from stdin via -, so any command that emits URLs feeds a walk:

bili search lofi -o url | bili discover - --depth 1

When an edge is gated

Most read endpoints answer anonymous callers, but a few (creator catalogues in particular) are gated by bilibili's anti-bot for some IPs. The walk treats the two cases differently:

  • A seed that cannot be fetched fails the walk, like any bad id.
  • An edge that fails deeper in the walk becomes a one-line note on stderr and the walk carries on with the other edges.

Supplying a logged-in --cookie or BILI_COOKIE widens what a walk can reach; -q silences the notes.

discover or crawl?

Both walk the graph from seeds, but they are built for different jobs:

  • discover streams one record per node to stdout. It is for exploring, piping, and rendering in any output format. To keep a walk, redirect it: bili discover BV… --depth 2 -o jsonl > graph.jsonl.
  • crawl writes one JSONL file per record type into a directory, and pulls attached data like comments and danmaku. It is for building a split dataset on disk. See crawling.

Reach for discover to look around; reach for crawl to bulk-dump.

3 min read