Frontmatter MCP

# 9. Use ruri-v3-30m for Japanese text embedding Date: 2025-12-05 ## Status Accepted ## Context Semantic search requires an embedding model to convert text into vectors. The primary use case is searching Japanese Markdown notes (e.g., Obsidian vault). Requirements: - Japanese language support with high quality - Lightweight for local execution on consumer hardware (e.g., M1 Mac) - No external API calls (zero cost, offline capable) Models considered: | Model | Parameters | JMTEB Score | Notes | |-------|------------|-------------|-------| | cl-nagoya/ruri-v3-30m | 30M | 72.95 | Japanese-specialized | | multilingual-e5-small | 118M | 67.38 | Multilingual | | multilingual-e5-base | 278M | 70.53 | Multilingual | ## Decision Adopted cl-nagoya/ruri-v3-30m as the default embedding model. Key factors: - Highest JMTEB score (72.95) among lightweight models - Smallest parameter count (30M) enables fast inference - 256-dimensional output balances quality and storage efficiency - Japanese-specialized training data The model is configurable via `FRONTMATTER_EMBEDDING_MODEL` environment variable for users who prefer different models. ## Consequences Benefits: - Fast embedding generation even on CPU - Small model download size (~120MB) - High quality Japanese text understanding Trade-offs: - Optimized for Japanese; may underperform for other languages - Smaller dimension (256) compared to larger models (768+) may lose some nuance - Users with multilingual notes may need to switch models