vidistill 0.3.3

Video content distillation CLI tool

vidistill

A toolset to download, parse and analyse social-media video content.

Install

uv sync
uv run playwright install chromium   # one-time: downloads ~150 MB browser binary

playwright install chromium is required by the collect subcommand, which drives a headless browser against Bilibili pages.

Environment

Copy .env.example → .env and fill in:

• OPENROUTER_API_KEY — LLM provider for analyse.
• VOLC_ACCESS_KEY_ID / VOLC_SECRET_ACCESS_KEY / VOLC_TOS_BUCKET — Volcengine TOS + 豆包 ASR credentials used by the default transcriber.

CLI

vidistill analyse <input>  [--output-dir …] [--model …] [--transcriber …] [--diarize]
vidistill collect <page-path> [--rule …] [--num …] [--sort …] [--max-scrolls …] [--headful]

vidistill analyse

Runs the full pipeline on a single video: download → audio extraction → transcription → LLM analysis → markdown report.

vidistill analyse https://www.bilibili.com/video/BV1xx411c7mu

vidistill collect

Scrape a Bilibili listing page, filter videos by a rule expression, and write the top-N matches as JSON. Supported page types:

Type	URL shape
homepage	www.bilibili.com/ (the recommended feed)
user	space.bilibili.com/<uid>
search	search.bilibili.com/all?keyword=…
channel	www.bilibili.com/c/<slug>
popular	www.bilibili.com/v/popular/<column> (all, weekly, …)
topic	www.bilibili.com/v/topic/detail?topic_id=<id>

vidistill collect https://space.bilibili.com/12345 \
    --rule='play>10000 and like>=500' --num=20

vidistill collect "https://search.bilibili.com/all?keyword=LLM" \
    -r 'play>50000 and duration<=600' -n 10 --sort=play

Rule DSL (--rule) supports comparisons (== != > >= < <=), boolean ops (and, or, not), parentheses, and the following fields: play, like, coin, favorite (alias star), share, danmaku, comment (alias reply), duration (seconds), publish_days_ago, title, author. Missing fields evaluate the whole rule to False for that video.

Output lands in <output-dir>/collect_<page-type>_<identifier>_<ts>.json.

Bilibili login & storage state

Bilibili's anti-bot layer rejects guest scraping. The --storage-state flag uses Playwright's native storage-state (cookies + localStorage) to persist login across runs. Default: default.storage_state in the current directory.

First run (no login yet):

vidistill collect https://space.bilibili.com/<uid> -n 10

If default.storage_state doesn't exist, a browser window opens. Log in to Bilibili there. When you close the browser, storage state is saved automatically and scraping proceeds. Subsequent runs reuse that state.

Explicit path:

vidistill collect <url> --storage-state ~/my-session.storage_state -n 10

Development

uv run pytest