Remote Input

Remote Input

Download remote content (arxiv papers, YouTube videos, Bilibili videos) to local storage and seamlessly integrate into the downstream pipeline.

What This Skill Does

This skill acts as a pre-processing layer for the pipeline:

1. Detects URL type (arxiv, YouTube, or Bilibili)
2. Downloads content to data/raw_inputs/remote/
3. Returns the local file path
4. Triggers the next pipeline stage

Supported URL Types

Note on merge failure: When video download produces separate video and audio files (merge failure), the downloader returns merge_failed: true with audio_path pointing to the audio file. The downstream routing switches to meeting-audio-grounding instead of meeting-video-grounding.

Directory Structure

data/raw_inputs/remote/
├── arxiv/
│   └── <paper_id>.pdf          # e.g., 2301.07041.pdf
├── youtube/
│   └── <video_title>.mp4       # e.g., "Introduction to Transformers.mp4"
├── bilibili/
│   └── <video_title>.mp4       # e.g., "教程视频.mp4"
└── metadata/
    └── <ground_id>.json        # Download metadata

Workflow

Step 1. Parse Input URL

Detect the URL type:

If URL contains "arxiv.org":
    → Use arxiv downloader
    → Ground ID = arxiv paper ID (e.g., "2301.07041")

If URL contains "youtube.com" or "youtu.be":
    → Use YouTube downloader
    → Ground ID = sanitized video title or video ID

If URL contains "bilibili.com" or "b23.tv":
    → Use Bilibili downloader
    → Ground ID = BV ID (e.g., "BV1xx411c7JZ")

Step 2. Download Content

For arXiv Papers

python .cursor/skills/remote-input/scripts/download_arxiv.py download "<url_or_id>" \
    --dir data/raw_inputs/remote/arxiv

Output:

{
  "success": true,
  "path": "data/raw_inputs/remote/arxiv/2301.07041.pdf",
  "paper_id": "2301.07041",
  "title": "Attention Is All You Need",
  "authors": ["Ashish Vaswani", ...],
  "abstract": "The dominant sequence transduction models...",
  "size_kb": 1024
}

For YouTube Videos

python .cursor/skills/remote-input/scripts/download_video.py "<url>" \
    -o data/raw_inputs/remote/youtube \
    -q 720p

Output (normal - video with audio):

{
  "success": true,
  "path": "data/raw_inputs/remote/youtube/Video Title.mp4",
  "title": "Video Title",
  "video_id": "dQw4w9WgXcQ",
  "source": "youtube",
  "duration": 213,
  "uploader": "Rick Astley",
  "audio_path": null,
  "merge_failed": false
}

Output (merge failure - separate audio file):

{
  "success": true,
  "path": "data/raw_inputs/remote/youtube/Video Title.f136.mp4",
  "title": "Video Title",
  "video_id": "dQw4w9WgXcQ",
  "source": "youtube",
  "duration": 213,
  "uploader": "Rick Astley",
  "audio_path": "data/raw_inputs/remote/youtube/Video Title.f251.webm",
  "merge_failed": true
}

For Bilibili Videos

python .cursor/skills/remote-input/scripts/download_video.py "<url>" \
    -o data/raw_inputs/remote/bilibili \
    -q 720p \
    -s  # 可选：下载字幕

Output (normal - video with audio):

{
  "success": true,
  "path": "data/raw_inputs/remote/bilibili/视频标题.mp4",
  "title": "视频标题",
  "video_id": "BV1xx411c7JZ",
  "source": "bilibili",
  "duration": 600,
  "uploader": "UP主名称",
  "audio_path": null,
  "merge_failed": false
}

Output (merge failure - separate audio file):

{
  "success": true,
  "path": "data/raw_inputs/remote/bilibili/视频标题.f136.mp4",
  "title": "视频标题",
  "video_id": "BV1xx411c7JZ",
  "source": "bilibili",
  "duration": 600,
  "uploader": "UP主名称",
  "audio_path": "data/raw_inputs/remote/bilibili/视频标题.f251.webm",
  "merge_failed": true
}

Note: Bilibili 视频需要登录 Cookie 才能下载高画质 (1080p+)。使用默认设置可下载 1080p 及以下画质。如需下载更高画质，请配置 --cookies-from-browser chrome 或提供 cookie 文件。

Step 3. Write Metadata

Save download metadata to:

data/raw_inputs/remote/metadata/<ground_id>.json

Example:

{
  "ground_id": "arxiv_2301.07041",
  "source_type": "arxiv",
  "source_url": "https://arxiv.org/abs/2301.07041",
  "downloaded_path": "data/raw_inputs/remote/arxiv/2301.07041.pdf",
  "downloaded_at": "2024-01-15T10:30:00Z",
  "metadata": {
    "title": "Attention Is All You Need",
    "authors": ["Ashish Vaswani", ...]
  }
}

Example (Bilibili):

{
  "ground_id": "bilibili_BV1xx411c7JZ",
  "source_type": "bilibili",
  "source_url": "https://bilibili.com/video/BV1xx411c7JZ",
  "downloaded_path": "data/raw_inputs/remote/bilibili/视频标题.mp4",
  "downloaded_at": "2024-01-15T10:30:00Z",
  "metadata": {
    "title": "视频标题",
    "uploader": "UP主名称"
  }
}

Step 4. Return Local Path

Return the local file path for downstream pipeline integration:

• For one-report: Pass the local path as input_path
• For input-router: The router will detect .pdf or .mp4 extension

Usage Examples

Example 1: arXiv Paper

Input: https://arxiv.org/abs/2301.07041

Workflow:
1. Download PDF to: data/raw_inputs/remote/arxiv/2301.07041.pdf
2. Write metadata: data/raw_inputs/remote/metadata/arxiv_2301.07041.json
3. Return: data/raw_inputs/remote/arxiv/2301.07041.pdf

Next: Pass to input-router → document-grounding → ...

Example 2: YouTube Video

Input: https://www.youtube.com/watch?v=dQw4w9WgXcQ

Workflow:
1. Download video to: data/raw_inputs/remote/youtube/Rick Astley - Never Gonna Give You Up.mp4
2. Write metadata: data/raw_inputs/remote/metadata/youtube_dQw4w9WgXcQ.json
3. Return: data/raw_inputs/remote/youtube/Rick Astley - Never Gonna Give You Up.mp4

Next: Pass to input-router → meeting-video-grounding → ...

Example 2b: Bilibili Video

Input: https://bilibili.com/video/BV1xx411c7JZ

Workflow:
1. Download video to: data/raw_inputs/remote/bilibili/视频标题.mp4
2. Write metadata: data/raw_inputs/remote/metadata/bilibili_BV1xx411c7JZ.json
3. Return: data/raw_inputs/remote/bilibili/视频标题.mp4

Next: Pass to input-router → meeting-video-grounding → ...

Supported Bilibili URL formats:

• https://bilibili.com/video/BV1xx411c7JZ (BV号)
• https://www.bilibili.com/video/av12345678 (AV号)
• https://b23.tv/abc123 (短链接)

Example 3: Via one-report Skill

Use the one-report skill with a remote URL:

Input:
- input_path: https://arxiv.org/abs/2301.07041
- output_formats: md,pdf

The one-report skill will:
1. Detect URL input (not local file)
2. Invoke remote-input to download
3. Continue pipeline with local file path

Integration Points

Integration with one-report

When input_path is a URL:

1. Detect URL pattern
2. Invoke remote-input skill
3. Use returned local path for downstream pipeline

Integration with input-router

Update routing table to support URLs:

Error Handling

Merge Failure Handling

When video download produces separate video and audio files (merge failure):

Detection

The downloader uses ffprobe to check if the main video file contains an audio track:

• If no audio track detected → triggers merge attempt
• If merge fails → marks as merge_failed: true

Auto-Recovery Flow

1. Download produces: video.mp4 (no audio) + audio.webm
2. Detect: video.mp4 lacks audio
3. Attempt: ffmpeg merge video.mp4 + audio.webm → video.mp4
4. If merge succeeds → use video.mp4 (merge_failed: false)
5. If merge fails → use audio.webm directly (merge_failed: true)

Downstream Impact

When merge_failed: true, the downloader returns:

{
  "success": true,
  "path": "video.mp4",
  "audio_path": "audio.webm",
  "merge_failed": true
}

The pipeline should:

1. Use audio_path for transcription instead of video
2. Skip video-only processing
3. Continue with meeting-audio-grounding using the audio file

Supported Audio Formats for Transcription

WhisperX (used by audio_structuring) supports:

• .mp3, .wav, .m4a, .aac, .ogg
• .webm (also supported via ffmpeg backend)

The audio file at audio_path can be directly passed to meeting-audio-grounding.

Ground ID Generation

Quality Settings

Note: For Bilibili, default settings download up to 1080p from domestic servers. Higher quality (4K/8K) requires login cookies configuration.

For YouTube

Default quality: 720p (recommended for most use cases)

For Bilibili

Bilibili Login Configuration (Optional)

For higher quality downloads:

# Use browser cookies
--cookies-from-browser chrome

# Or provide cookie file
--cookies /path/to/cookies.txt

Audio-Only Option

For YouTube videos that are primarily audio (podcasts, lectures):

python scripts/download_video.py "<url>" -a

Downloads audio as MP3 to the youtube directory.