RAG grounds the assistant in writings, research, automations, and project data stored in this site's knowledge base.
In my daily workflow
- I keep source content structured: headings, metadata, publish status.
- Embeddings refresh when content changes — tied to admin publish actions.
- Retrieval uses hybrid search where keyword anchors matter (titles, slugs).
- Answers must cite chunks; no citation means downgrade or refuse.
How it makes me work smarter
RAG lets a small model punch above its weight on domain questions. Visitors ask about my stack or a specific automation; retrieval pulls the right page instead of relying on training data. I spend effort on chunk quality and metadata, not bigger models.
My setup
- Supabase pgvector for embeddings
- Chunking by semantic sections with overlap
- Metadata filters: status=published, content type
- Reranker optional for high-stakes answers
On this portfolio
The knowledge base powers public pages and feeds the AI assistant's retrieval layer. Admin draft/archived status prevents unpublished notes from entering the index — the same status dropdown I use daily in `/admin/content`.
