Video Analyzer

Use Cases

• Users request to analyze, understand, or describe a video
• Users provide a video file path or URL and want to know the video content
• Users need to ask questions about the video

Configuration

Environment Variables

# VolcEngine (Doubao)
export ARK_API_KEY="your-api-key"

# OpenAI
export OPENAI_API_KEY="your-api-key"

Model Configuration

scripts/models.json

• base_url

  — API endpoint

• api_key_env

  — Name of the environment variable for reading the API Key

• model

  — Model ID

• api_type

  —

  responses

  or

  chat_completions

• supports_video

  — Whether native video input is supported

Workflow

1. Confirm Video Source: Obtain the video path or URL provided by the user.
2. Confirm Analysis Requirements: Clarify what the user wants to know (e.g., content summary, question answering, scene description, etc.). If

   $ARGUMENTS

   is not empty, use it as the analysis prompt.
3. Select Model: By default, use the

   default_model

   in

   models.json

   ; users can also specify a model.
4. Execute Analysis: Run the script (execute in the

   scripts/

   directory):
   bash

   uv run analyze.py --video <video path or URL> --prompt "<analysis prompt>"

   Optional parameters:

   • --model <name>

     — Specify a model (corresponds to the key in models.json)

   • --frames <number>

     — Number of frames to extract (default: 10)

   • --max-size <pixels>

     — Maximum side length of frames (default: 720)
5. Display Results: Present the analysis results returned by the model to the user.

CLI Reference

# Local video
uv run analyze.py --video /path/to/video.mp4 --prompt "Describe the video content"

# Direct online video URL
uv run analyze.py --video https://example.com/video.mp4 --prompt "Analyze the video"

# Video website URL (YouTube, Bilibili, etc.)
uv run analyze.py --video https://www.youtube.com/watch?v=xxxxx --prompt "Summarize the video"

# Specify model and number of frames
uv run analyze.py --video video.mp4 --model doubao-vision --frames 20 --prompt "Analyze"

Notes

• Downloading video from website URLs depends on

  yt-dlp

  , which is automatically installed as a Python dependency
• In frame extraction mode, more frames lead to more detailed analysis but higher API call costs
• Downloading large video files may take a long time, please be patient