v0.0.2 - Semantic Transcript Search

✨ New Features

Semantic Transcript Search

Search YouTube video transcripts using natural language queries!

New MCP Tools:

• warmup_semantic_search - Pre-load embedding model (~270MB) to avoid timeout on first search
• semantic_search_transcripts - Search transcripts with auto-indexing
• index_channel_transcripts - Optional bulk pre-warming for faster searches

Example

# Search Vimjoyer's channel for NixOS topics
results = await semantic_search_transcripts(
    query="how to configure garbage collection",
    channel_ids=["UCuAXFkgsw1L7xaCfnd5JJOw"],
)

# Returns timestamped URLs for direct playback
print(results["results"][0]["timestamp_url"])
# https://www.youtube.com/watch?v=DnA4xNTrrqY&t=231

🔧 Technical Details

• Embeddings: Nomic Embed Text v1.5 with Matryoshka dimensionality (512 dims)
• Vector Store: ChromaDB with optimized HNSW settings
• Chunking: Token-based (256 tokens) with timestamp preservation
• Tests: 343 passing (was 178)

📦 Installation

pip install yt-api-mcp==0.0.2

# Or with uvx
uvx yt-api-mcp stdio

Full Changelog

See CHANGELOG.md