RAG

End-to-end Retrieval-Augmented Generation on NotNull Online. Hybrid retrieval combines MongoDB Atlas Vector Search (semantic) and Atlas Search (keyword), with a compact published catalog for grounding context. Pipeline covers query embedding, retrieval merge, prompt construction, grounded validation and fallback when evidence is weak, source citation derivation, and SSE token streaming to the Angular client. Design constraint: answers must stay grounded in knowledge_items — no fabricated projects, employers, or skills when retrieval finds no supporting evidence. Provider access goes through domain AI ports so completions and embeddings stay swappable without rewriting the RAG use case.