Implements local RAG (Retrieval Augmented Generation) using pre-computed
CLIP embeddings for visual prompt enhancement:
- core/compose/rag/ module with index loading, retrieval, and build script
- CLIP-based image embeddings (openai/clip-vit-base-patch32)
- Pure numpy cosine similarity search (no external services)
- Category-aware retrieval mapped to canonical categories
- Pre-built index with 613 prompts from scraped reference data
New Compose node parameters:
- use_rag: Enable visual RAG vocabulary enhancement
- rag_threshold: Minimum similarity threshold (0.5-0.95)
RAG tokens integrated as new TokenSource with 0.85 weight.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
