feat(conversation-memory): upgrade to EmbeddingGemma for better multilingual support

[baleen37]

Summary

Upgrades conversation-memory plugin from all-MiniLM-L6-v2 to EmbeddingGemma-300M for significantly better multilingual (especially Korean) support.

Key Changes

• Embedding Model: Switch to onnx-community/embeddinggemma-300m-ONNX with Q4 quantization
• Performance: 51% better Korean performance (83.86 vs 55.4 MRR@10)
• Dimensions: 768-dim embeddings (vs 384) for better semantic representation
• Languages: 100+ language support with MTEB Multilingual score of 60.62
• Package: Migrated from @xenova/transformers to @huggingface/transformers v3

Breaking Changes

⚠️ Database must be recreated due to embedding dimension mismatch (384 → 768).

Migration guide added to README with step-by-step instructions.

Technical Details

• Model size: ~197MB (Q4 quantization)
• Memory usage: < 200MB RAM
• Task prefix: "title: none | text: ..." (automatically applied)
• Cache directory: ./.cache for model storage

Test Plan

• Pre-commit hooks pass
• Manual testing: conversation indexing
• Manual testing: semantic search
• Verify model downloads and caches correctly

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Updated embedding model with enhanced 768-dimensional vector support for improved search relevance
  • v2.0 includes multilingual capabilities and optimized memory efficiency

• Documentation

  • Added comprehensive migration guide for upgrading from v1.x to v2.0
  • Expanded documentation with architecture details, updated examples, and storage structure information

[coderabbitai]

Caution

Review failed

The pull request is closed.

📝 Walkthrough

Walkthrough

The conversation-memory plugin is upgraded to use the HuggingFace transformers library with the EmbeddingGemma model (768-dimensional embeddings with Q4 quantization) replacing Xenova's 384-dimensional variant. Documentation is substantially expanded with migration guidance and architecture details.

Changes

Cohort / File(s)	Summary
Documentation & Migration plugins/conversation-memory/README.md	Expanded with upgrading from v1.x migration steps, What's New in v2.0 section with model metrics, architecture notes for EmbeddingGemma, and updated dependency references (huggingface/transformers).
Dependencies plugins/conversation-memory/package.json	Swapped transformer provider from @xenova/transformers to @huggingface/transformers.
Vector Schema & Embedding Pipeline plugins/conversation-memory/src/core/db.ts, plugins/conversation-memory/src/core/embeddings.ts	Updated vector embedding dimension from 384 to 768 in database schema. Migrated embedding model to EmbeddingGemma (ONNX) with Q4 quantization, added mandatory task prefix format ("title: none | text: ..."), and expanded token truncation limit from 2000 to 8000.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~22 minutes

Suggested labels

enhancement

Poem

🐰 Gemma hops in with dimensions galore,
768 embeddings, not 384 anymore!
Task prefixes align, transformers renewed,
The memory plugin's embedding mood!

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch feat/conversation-memory-embeddinggemma

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}