[PRD] RAG / Design Knowledge Base

[espetro]

Summary

Add memory to the AI pipeline. Currently, each generation starts fresh with no memory of previous work. RAG captures design patterns, user preferences, brand guidelines, and past iterations to improve generation quality and reduce iteration count by 30-40%.

Deliverables

• SQLite-vss integration for vector similarity search
• Knowledge entry CRUD (patterns, preferences, brand guidelines)
• Implicit knowledge capture from critique output
• Context injection into all 5 pipeline stages
• Knowledge panel UI (Patterns, Preferences, Brand tabs)
• Export/import knowledge bases

Effort

• MVP (Weeks 1-4): ~20 story points — SQLite-vss, basic retrieval, implicit capture
• v1.0 (Weeks 5-10): ~40 story points — Full pipeline retrieval, explicit UI, brand guidelines
• v2.0 (Future): ~55 story points — Pattern insights, similarity explorer, multi-device sync

Timeline

Post-v1.0 — Requires SQLite persistence layer (still scaffolding). Ship after core product is stable.

References

• 📄 PRD: docs/roadmap/rag-design-knowledge-base.md
• 🏗️ Architecture: Layered repository pattern, SQLite-vss for vectors, nomic-embed-text-v1.5 for local embeddings
• 🔗 Depends on: packages/db (SQLite+Drizzle), comment/revision system (exists)