YouTube Summarizer

Turns YouTube videos into structured Markdown reports with YAML frontmatter. Uses a map-reduce LLM pipeline: chunks the transcript, extracts key facts in parallel, then synthesizes into a final article.

Setup

1. Dependencies

   pip install -r requirements.txt
   

2. Install yt-dlp (binary, not pip)

   brew install yt-dlp   # macOS
   # or: pip install yt-dlp
   

3. Install a local LLM server (e.g. Ollama, vLLM) and configure config.toml:

   [llm]
   endpoint = "http://localhost:9931/v1"
   model_name = "your-model-name"
   

4. Add channels to [channels] in config.toml (used by batch_runner.py).

Usage

Single video:

python summarizer.py "https://www.youtube.com/watch?v=VIDEO_ID"

All configured channels:

python batch_runner.py

Output

Summaries are written to outputs/summaries/ as:

{video_id}_{slug_title}.md

Each file has YAML frontmatter with metadata (title, channel, published date, duration, tags, speakers, key takeaways) followed by the markdown report.

How it works

1. Extract — grabs transcript (via youtube-transcript-api or yt-dlp) and video metadata (scrapes YouTube)
2. Chunk — splits the transcript into overlapping chunks
3. Map — sends each chunk to the LLM in parallel to extract key facts
4. Reduce — synthesizes all extracted facts into one flowing report

Config

config.toml sections:

• [llm] — endpoint, model, timeout, retries
• [processing] — chunk size, overlap, max parallel workers
• [channels] — channel name → homepage URL (for batch runs)
• [yt_dlp] — retries, socket timeout, geo-bypass, extractor args
• [paths] — output, summary, and transcript directories