A RAG system that's actually production-ready — three independent layers instead of one fragile pipeline.
Knowledge (PDF → chunks → embeddings → ChromaDB, with semantic topic metadata), retrieval (metadata pre-filter, then top-k vector search) and validation (confidence score per chunk, file + page citations assembled before the LLM call). Every layer is testable and replaceable.